Exterior differential systems

INTRODUCTION This book gives a treatment of exterior differential systems. It will include both the general theory and v...

Author: Bryant R.L. | Chern S.S. | Gardner R.B. | Goldschmidt H.H. | Griffiths P.A.

60 downloads 1282 Views 3MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

INTRODUCTION

This book gives a treatment of exterior differential systems. It will include both the general theory and various applications. An exterior differential system is a system of equations on a manifold defined by equating to zero a number of exterior differential forms. When all the forms are linear, it is called a pfaffian system. Our object is to study its integral manifolds, i.e., submanifolds satisfying all the equations of the system. A fundamental fact is that every equation implies the one obtained by exterior differentiation, so that the complete set of equations associated to an exterior differential system constitutes a differential ideal in the algebra of all smooth forms. Thus the theory is coordinatefree and computations typically have an algebraic character; however, even when coordinates are used in intermediate steps, the use of exterior algebra helps to efficiently guide the computations, and as a consequence the treatment adapts well to geometrical and physical problems. A system of partial differential equations, with any number of independent and dependent variables and involving partial derivatives of any order, can be written as an exterior differential system. In this case we are interested in integral manifolds on which certain coordinates remain independent. The corresponding notion in exterior differential systems is the independence condition: certain pfaffian forms remain linearly independent. Partial differential equations and exterior differential systems with an independence condition are essentially the same object. The latter, however, possess some advantages among which are the facts that the forms themselves often have a geometrical meaning, and that the symmetries of the exterior differential system are larger than those generated simply by changes of independent and dependent variables. Another advantage is that the coordinatefree treatment naturally leads to the intrinsic features of many systems of partial differential equations. It was Pfaff who pioneered the study of exterior differential systems by his formulation of the Pfaff problem in Pfaff [1814-15]. The exterior derivative of a pfaffian form, called the bilinear covariant, was introduced by Frobenius in 1877 and effi´ Cartan ciently used by Darboux in Darboux [1882]. In his book Cartan [1922], Elie introduced exterior differential forms of higher degree and their exterior derivatives. In 1904–08 he was led to the notion of a pfaffian system in involution through his work in generalizing the Maurer–Cartan forms in Lie groups to infinite Lie pseudogroups. The geometrical concepts introduced in this study apply to general exterior differential systems, as recognized by Goursat. An authoritative account was given in K¨ ahler [1934], culminating in an existence theorem now known as the Cartan–K¨ ahler theorem.

1

2

Introduction

On the side of partial differential equations the basic existence theorem is of course the Cauchy–Kowalewski theorem. More general existence theorems were given by Riquier [1910]. The use of differential operators in studying differential geometry has been traditional and has an extensive literature. Among our fundamental concepts are prolongation and involutivity. Intuitively the former is the classical way of adjoining the partial derivatives themselves as new variables, and taking as new equations those obtained by differentiating the old set, while the latter is the property that further prolongations will not give essentially new integrability conditions. Their precise definitions are more subtle, and will be given in Chapters VI and IV respectively. A linear pfaffian system in involution is a “well-behaved” system. This concept entered in correspondence between Cartan and Einstein [1979] on ´ relativity. While Elie Cartan proved that the Einstein field equations in general relativity based on distant parallelism form an involutive system, Einstein was at first suspicious of the notion. Later he understood it and expressed his satisfaction and appreciation. Cartan expressed the involutivity condition in terms of certain integers, known as Cartan’s test. In modern language this is a homological condition. In fact, Serre proved in 1963 that involutivity is equivalent to the vanishing of certain cohomology groups (see Guillemin–Sternberg [1964]). This makes it possible to use the powerful tool of commutative algebra. At the very beginning one notices the similarity between polynomials and differential operators. It turns out that this relationship goes much deeper, and the theory involves a mixture of both commutative and exterior algebra. A fundamental problem is whether a given differential system will, after a finite number of prolongations lead to an involutive system. Cartan attempted to answer this question, but it was Kuranishi [1957] who finally proved the Cartan–Kuranishi prolongation theorem. The main tool is homology theory. A slightly weaker version of the theorem will be proved in this book. We should however emphasize that differential systems not in involution are just as important. In fact, they are probably richer in content. For example, nongeneric conditions on a manifold such as isometric embedding in low codimension or the presence of additional geometric structures frequently are expressed by a non-involutive system. The last half of Cartan [1946] and Partie II of his “Œuvres Complètes” (Cartan [1953]) are full of “examples”, many of which are topics in their own right. An objective of this book is to call attention to these beautiful results, which have so far been largely ignored. As the results are coordinate-free, the theory applies well to global problems and to non-linear problems. A guiding problem in the theory is the equivalence problem: Given two sets of linear differential forms θi , θ∗j in the coordinates xk , x∗l respectively, 1 ≤ i, j, k, l ≤ n, both linearly independent, and given a Lie group G ⊂ GL(n, R). To find the conditions that there are functions x∗i = x∗i (x1 , . . . , xn ), such that θ∗j , after the substitution of these functions, differ from θi by a transformation of G. This gives rise to an exterior differential system. Cartan’s idea was

Introduction

3

to introduce the parameters of G as new variables, setting ωi = gji θj , ω∗i = gj∗i θ∗j , (gji ), (gj∗i ) ∈ G. j

Then in the product of the manifold with G, the condition becomes ωi = ω∗i , which is symmetrical in both sides. This means we should study the problem in a principal G-bundle and the formulation becomes global. The equivalence problem gives local Riemannian geometry when G = O(n) and the local invariants of an almost complex structure when n = 2m and G = GL(m, C). Similarly, it gives the local invariants of CR-geometry when n = 2m − 1 and G is a suitable subgroup of GL(m, C). We have stated the equivalence problem because of its importance; it will not be explicitly treated in this book; see Gardner [1989] for a modern exposition. The subject is so rich that a worker in the field is torn between the devil of the general theory and the angel of geometrical applications, which present all kinds of interesting phenomena. We have attempted to strike a balance. We will develop the general theory both from the standpoint of exterior differential systems and from that of partial differential equations. We will also give a large number of applications. A summary of contents follows: Chapter I gives a review of exterior algebra, with emphasis on results which are relevant to exterior differential systems. For those who like an intrinsic treatment it includes an introduction to jet bundles. Chapter II treats some simple exterior differential systems, particularly those which can be put in a normal form by a change of coordinates. They include completely integrable systems (Frobenius theorem) and the pfaffian equation. Cauchy characteristics for exterior differential systems come up naturally. Some arithmetic invariants are introduced for pfaffian systems. Even a pfaffian system of codimension 2, only partially treated in the last section, is a rich subject, with several interesting applications. Chapter III discusses the generation of integral manifolds through the solution of a succession of initial-value problems. Various basic concepts are introduced. The Cartan–K¨ ahler theorem is given as a generalization of the Cauchy–Kowalewsky theorem; the proof follows that of Kähler. As an application we give a proof of the isometric imbedding theorem of Cartan–Janet. Chapter IV introduces the important concepts of involution, linear differential systems, tableau and torsion. For linear pfaffian systems the condition of involution, as expressed by Cartan’s test, takes a simple form that is useful in computing examples. We also introduce the concept of prolongation, which will be more fully developed in Chapter VI. As one example we show that the high-dimensional Cauchy–Riemann equations are in involution. We also study the system of q partial differential equations of the second order for one function in n variables and find conditions for their involutivity. A geometrical application is made to the problem of isometric surfaces preserving the lines of curvature. It is an example of an over-determined system which, after several prolongations, leads to a simple and elegant result. In this example the effectiveness of exterior differential systems is manifest.

4

Introduction

Chapter V introduces the characteristic variety of a differential system. Intuitively its tangent spaces are hyperplanes of integral elements whose extension fails to be unique. It plays as important a role as the characteristics in classical partial differential equations. For a linear pfaffian system a definition can be given in terms of the symbol and the two agree in the absence of Cauchy characteristics. We discuss in detail the case of surfaces in E 3 and their Darboux frames, as this example illustrates many of the basic notions of exterior differential systems. Some properties of the characteristic variety are given. The deeper ones require the system to be involutive and the use of the complex characteristic variety. Their proofs rely on results of commutative algebra and are postponed to Chapter VIII. Chapter VI treats prolongation, another well-known process in the case of partial differential equations. The issue is whether any system with an independence condition (I, Ω) can be prolonged to an involutive system in a finite number of steps (Cartan–Kuranishi theorem). With our definition of prolongation we prove that the first prolongation of an involutive linear pfaffian system is involutive, a result that does not seem to appear in the literature. We establish a somewhat weaker version of the Cartan–Kuranishi theorem, thus giving in a sense a positive answer to the above question. As usual the general theory is illustrated by a number of examples. Chapter VII is devoted to some examples and applications. We give a classification of systems of first-order partial differential equations of two functions in two variables. Other examples include: triply orthogonal systems, finiteness of web rank, isometric imbedding and the characteristic variety. In Chapter VIII we study the algebra of a linear pfaffian system and its prolongations. The crucial information is contained in the tableau. Its properties are given by the Spencer cohomology groups or the Koszul homology groups, which are dual to each other. Involutive tableau is characterized by the vanishing of certain Spencer cohomology or Koszul homology groups. It is a remarkable coincidence that a regular integral flag and a quasi-regular graded SV -module represent essentially the same object. Homological algebra provides the tools to complete the proofs of the theorems stated in Chapters V and VI, and in particular the Cartan–Kuranishi theorem. As a consequence sheaf theory in commutative algebra and micro-local analysis in partial differential equations become parallel developments. Chapters IX and X give an introduction to the Spencer theory of over-determined systems of partial differential equations. While Cartan began his theory in the study of infinite pseudogroups, Spencer had a similar objective, viz., the study of the deformations of pseudogroup structures. His approach is more in the spirit of Lie, with a full use of modern concepts. We see in our account more emphasis on the general theory, although many examples are given. We hope that after the exposition in this book we come to realize that exterior differential systems and partial differential equations are one and the same subject. It is conceivable that different attires are needed for different purposes. This book grew through our efforts to work through and appreciate Partie II of Cartan’s “Œuvres Complètes” (Cartan [1953]), which we found to be full of interesting ideas and details. Hopefully our presentation will help the study of the original work, which we cannot replace. In fact, for readers who have gone through most of this book we propose the following problem as a final examination: Give a report on his famous five-variable paper, “Les systèmes de Pfaff à cinq variables et les équations aux dérivées partielles du second ordre” (Cartan [1910]).

Introduction

æ

5

6

I. Preliminaries

CHAPTER I PRELIMINARIES

In this chapter we will set up some notations and conventions in exterior algebra, give a description of the basic topic of the book, and introduce the language of jets which allows easy passage between partial differential equations and exterior differential systems. In particular we establish some basic results in exterior algebra Theorems 1.3, 1.5, and 1.7 which will be used in Chapter II.

§1. Review of Exterior Algebra. Let V be a real vector space of dimension n and V ∗ its dual space. An element x ∈ V is called a vector and an element ω ∈ V ∗ a covector. V and V ∗ have a pairing x, ω, x ∈ V, ω ∈ V ∗ , which is a real number and is linear in each of the arguments, x, ω. Over V there is the exterior algebra, which is a graded algebra: Λ(V ) = Λ0 (V ) ⊕ Λ1 (V ) ⊕ · · · ⊕ Λn (V ), with Λ0 (V ) = R,

Λ1 (V ) = V.

An element ξ ∈ Λ(V ) can be written in a unique way as ξ = ξ0 + ξ1 + · · · + ξn , where ξp ∈ Λp (V ) is called the p-th component of ξ. An element ξ = ξp ∈ Λp (V ) is called homogeneous of degree p or a multivector of dimension p. Multiplication in Λ(V ) will be denoted by the wedge sign: ∧. It is associative, distributive, but not commutative. Instead it satisfies the relation ξ ∧ η = (−1)pq η ∧ ξ,

ξ ∈ Λp (V ),

η ∈ Λq (V ).

The multivector ξ is called decomposable, if it can be written as a monomial (1)

ξ = x1 ∧ · · · ∧ xp ,

We have the following fundamental fact:

xi ∈ V.

§1. Review of Exterior Algebra

7

Proposition 1.1 (Criterion on linear dependence). The vectors x1 , . . . , xp are linearly dependent if and only if x1 ∧ · · · ∧ xp = 0. If the decomposable multivector ξ in (1) is not zero, then x1 , . . . , xp are linearly independent and span a linear subspace W of dimension p in V . This space can be described by (2)

W = {x ∈ V | x ∧ ξ = 0}.

Let x1 , . . . , xp be another base in W . Then ξ = x1 ∧ · · · ∧ xp is a (non-zero) multiple of ξ. We will call ξ, defined up to a constant factor, the Grassmann coordinate vector of W , and write (3)

[ξ] = W,

the bracket indicating the class of coordinate vectors differing from each other by a non-zero factor. In the same way there is over V ∗ the exterior algebra Λ(V ∗ ) = Λ0 (V ∗ ) ⊕ Λ1 (V ∗ ) ⊕ · · · ⊕ Λn (V ∗ ), Λ0 (V ∗ ) = R, Λ1 (V ∗ ) = V ∗ . An element of Λp (V ∗ ) is called a form of degree p or simply a p-form. Let ei be a base of V and ωk its dual base, so that ei , ωk = δik ,

1 ≤ i, k ≤ n.

Then an element ξ ∈ Λp (V ) can be written (4)

ξ = 1/p!

ai1 ...ip ei1 ∧ · · · ∧ eip

and an element α ∈ Λp (V ∗ ) as (5)

α = 1/p!

bi1...ip ωi1 ∧ · · · ∧ ωip .

In (4) and (5) the coefficients ai1...ip and bi1...ip are supposed to be anti-symmetric in any two of their indices, so that they are well defined. It follows from (4) that any multivector is a linear combination of decomposable multivectors. For our applications it is important to establish the explicit duality or pairing of Λ(V ) and Λ(V ∗ ). We require that Λp (V ) and Λq (V ∗ ), p = q, annihilate each other. It therefore suffices to define the pairing of Λp (V ) and Λp (V ∗ ). Since, by the above remark, any multivector is a linear combination of decomposable multivectors, it suffices to have the pairing of ξ = x1 ∧ · · · ∧ xp ,

xi ∈ V,

8

I. Preliminaries

and α = ω1 ∧ · · · ∧ ωp ,

ωi ∈ V ∗ .

We will define ξ, α = det(xi , ωk ),

(6)

1 ≤ i, k ≤ p.

It can be immediately verified that this definition is meaningful, i.e., if ξ (resp. α) is expressed in a different way as a product of vectors (resp. covectors), the right-hand side of (6) remains unchanged. In terms of the expressions (4) and (5) the pairing is given by (7) ξ, α = 1/p! ai1 ...ip bi1 ...ip . This is proved by observing that the right-hand side of (7) is linear in the arguments ξ and α and that the right-hand sides of both (6) and (7) are equal when ξ and α are products of the elements of the dual bases. An endomorphism f of the additive structure of Λ(V ) is called a derivation of degree k if it satisfies the conditions: (i) f : Λp V → Λp+k V , 0 ≤ p ≤ n (ii) f(ξ ∧ η) = f(ξ) ∧ η + (−1)kp ξ ∧ f(η) for ξ ∈ Λp V , η ∈ ΛV . A derivation of degree −1 is also called an anti-derivation. Given ξ ∈ V , we define the exterior product e(ξ) : Λ(V ) → Λ(V ) by e(ξ)η = ξ ∧ η

η ∈ Λ(V ).

The adjoint operator of e(ξ), ξ

: Λ(V ∗ ) → Λ(V ∗ )

is called the interior product, and is defined by the relation η, ξ

α = e(ξ)η, α

η ∈ Λ(V ), α ∈ Λ(V ∗ ).

The following result is easily proved: Proposition 1.2. If x ∈ V , then x

is an anti-derivation.

Notice that e(x) is neither a derivation nor an anti-derivation. Definition. A subring I ⊂ Λ(V ∗ ) is called an ideal, if: a) α ∈ I implies α ∧ β ∈ I for all β ∈ Λ(V ∗ ); b) α ∈ I implies that all its components in Λ(V ∗ ) are contained in I. A subring satisfying the second condition is called homogeneous. As a consequence of a) and b) we conclude that α ∈ I implies β ∧ α ∈ I for all β ∈ Λ(V ∗ ). Thus all our ideals are homogeneous and two-sided. Given an ideal I ⊂ Λ(V ∗ ), we wish to determine the smallest subspace W ∗ ⊂ V ∗ such that I is generated, as an ideal, by a set S of elements of Λ(W ∗ ). An element of


9

I is then a sum of elements of the form σ ∧β, σ ∈ S, β ∈ Λ(V ∗ ). If x ∈ W = (W ∗ )⊥ , we have, since the interior product x is an anti-derivation, x x

σ = 0,

(σ ∧ β) = ±σ ∧ (x

β) ∈ I.

Therefore we define A(I) = {x ∈ V | x

I ⊂ I},

where the last condition means that x α ∈ I, for all α ∈ I. A(I) is clearly a subspace of V . It will play later an important role in differential systems, for which reason we will call it the Cauchy characteristic space of I. Its annihilator C(I) = A(I)⊥ ⊂ V ∗ will be called the retracting subspace of I. Theorem 1.3 (Retraction theorem). Let I be an ideal of Λ(V ∗ ). Its retracting subspace C(I) is the smallest subspace of V ∗ such that Λ(C(I)) contains a set S of elements generating I as an ideal. The set S also generates an ideal J in Λ(C(I)), to be called a retracting ideal of I. There exists a mapping ∆ : Λ(V ∗ ) → Λ(C(I)) of graded algebras such that ∆(I) = J. Proof. Suppose W ∗ ⊂ V ∗ be a subspace such that Λ(W ∗ ) contains a set S of elements which generate I as an ideal. By the above discussion, if x ∈ W = (W ∗ )⊥ , we have x I ⊂ I. It follows that W ⊂ A(I), and consequently, C(I) = (A(I))⊥ ⊂ W ∗. We now choose a complementary space B of C(I) in V ∗ , so that V ∗ = B ⊕ C(I). Let ωi , 1 ≤ i ≤ n, be a base in V ∗ with ω1 , . . . , ωk ∈ B,

ωk+1 , . . . , ωn ∈ C(I).

Its dual base ei , 1 ≤ i ≤ n, then has the property that A(I) = {e1 , . . . , ek }. We define hj : Λ(V ∗ ) → Λ(V ∗ ), 1 ≤ j ≤ k, by hj (α) = α − ωj ∧ (ej

α),

α ∈ Λ(V ∗ ).

It is easy to verify that hj (α ∧ β) = hj (α) ∧ hj (β), so that each hj is a mapping of graded algebras. The same is therefore true of the composition ∆ = hk ◦ · · · ◦ h1 . Since ej ∈ A(I), 1 ≤ j ≤ k, we have hj (I) ⊂ I, from which we get ∆(I) ⊂ I. Clearly we have the restrictions ∆|B = 0, ∆|C(I) = Id. Since ∆ is a mapping of graded algebras, this implies that Λ(C(I)) is the image of ∆.

10

I. Preliminaries

It remains to construct the set S in Λ(C(I)) which generates I. This is done by induction on the degrees of the elements of I. Let Ip be the set of elements of I of degree p. To exclude the trivial case that I = Λ(V ∗ ) itself, we suppose I0 = ∅. Using this assumption we have, by the definition of A(I), x α = x, α = 0, x ∈ A(I), α ∈ I1 . It follows that A(I) ⊂ I1⊥ of I1 ⊂ C(I). To apply induction suppose that I1 , . . . , Ip−1 are generated by elements of Λ(C(I)). Denote by Jp−1 the ideal generated by them. Recall that hj , 1 ≤ j ≤ k, are mappings of graded algebras and induce the identity mapping on C(I). They therefore leave Jp−1 invariant. By the definition of h1 we have h1 (α) − α ∈ Jp−1 . Applying h2 , . . . , hk successively, we get ∆(α) − α ∈ Jp−1 . By replacing α by ∆(α) as a generator of I, we complete the induction. We wish to make some applications of the retraction theorem (Theorem 1.3). First we recall that, dualizing (2) and (3), the Grassmann coordinate vector [α] of a subspace W ∗ ⊂ V ∗ of dimension p is a non-zero decomposable p-covector such that W ∗ = {ω ∈ V ∗ | ω ∧ α = 0}. This notion can be extended to any p-form α, decomposable or not, by defining Lα = {ω ∈ V ∗ | ω ∧ α = 0}. Lα will be called the space of linear divisors of α, because of the property given in the following theorem: Proposition 1.4. Given a p-form α, let ω1 , . . . , ωq be a base for Lα . Then α may be written in the form α = ω1 ∧ · · · ∧ ωq ∧ π, with π ∈ Λp−q (V ∗ ). Proof. Take first the case q = 1. We can suppose ω1 to be a base element of V ∗ . By the expression (5) we can write α = ω 1 ∧ π + α1 , where α1 does not involve ω1 . The hypothesis ω1 ∧ α = 0 implies α1 = 0, so that the statement is true. The general case follows by induction on q. Theorem 1.5. Let I be an ideal generated by the linearly independent elements ω1 , . . . , ωs ∈ V ∗ and the 2-form Ω ∈ Λ2 (V ∗ ). Let p be the smallest integer such that (8)

Ωp+1 ∧ ω1 ∧ · · · ∧ ωs = 0.

Then the retracting space C(I) is of dimension 2p + s and has the Grassmann coordinate vector Ωp ∧ ω 1 ∧ · · · ∧ ω s . Proof. Consider first the case s = 0. An element of I is a linear combination of Ω, Ω2 , . . . , Ωp = 0. Hence by Theorem 1.3, we have Ω ∈ Λ(C(I)), and Ωp ∈ Λ2p(C(I)). The latter implies dim C(I) ≥ 2p.


Let

11

f : V → V∗

be the linear map defined by f(x) = x

Ω,

x ∈ V.

Since I does not contain a linear form, we have x

Ω = 0 if and only if x ∈ A(I) = C(I)⊥ .

This proves ker f = A(I), so that dim ker f = dim A(I) ≤ n − 2p.

(9)

On the other hand, the equation (8) gives for s = 0, x

Ωp+1 = (p + 1)(x

Ω) ∧ Ωp = 0.

Hence the space of linear divisors of Ωp contains the image of f. Since Ωp is of degree 2p and has at most 2p linear divisors, we have dim im f ≤ 2p.

(10)

Now it is an elementary fact that dim ker f + dim im f = n. Therefore the equality signs hold in both (9) and (10). In particular, we have dim C(I) = 2p and Λ2p(C(I)) is of dimension one, with Ωp as a base element, which is thus a Grassmann coordinate vector of C(I). In the general case, let W ∗ = {ω1 , . . . , ωs } be the space spanned by the ω’s. Then W = (W ∗ )⊥ ⊂ V and the quotient space V ∗ /W ∗ have a pairing induced by that of V and V ∗ , and are dual vector spaces. We have 0 = Ωp ∧ ω1 ∧ · · · ∧ ωs ∈ Λ2p+s (C(I)), so that dim C(I) ≥ 2p + s. Consider the linear map f

W − → V∗ − → V ∗ /W ∗ , π

where π is the projection (into a quotient space) and f is defined by f(x) = x

Ω, x ∈ W.

As above, we wish to find upper bounds for the dimensions of the kernel and image of f = π ◦ f. The sum of these dimensions is dim ker f + dim im f = n − s. A generalization of the above argument gives dim ker f ≤ n − 2p − s dim im f ≤ 2p Hence the equality signs hold everywhere and the theorem follows as before.

12

I. Preliminaries

Proposition 1.6. Let ω1 , . . . , ωs , π be linearly independent elements of V ∗ and Ω ∈ Λ2 V ∗ ; then (11)

Ωp ∧ ω 1 ∧ · · · ∧ ω s ∧ π = 0

implies Ωp+1 ∧ ω1 ∧ · · · ∧ ωs = 0. Proof. Let {π} denote the one dimensional space spanned by π and let W ∗ denote a complement in V ∗ of {π} which contains ω1 , . . . , ωs . Then there exist α ∈ Λ2 W ∗ , β ∈ W ∗ , uniquely determined, such that Ω = α + β ∧ π. It follows that Ωp = αp + pαp−1 ∧ β ∧ π and the hypothesis (11) implies αp ∧ ω1 ∧ · · · ∧ ωs ∧ π = 0. Since αp ∧ ω1 ∧ · · · ∧ ωs ∈ Λ(W ∗ ), we must have αp ∧ ω1 ∧ · · · ∧ ωs = 0. The conclusion now follows since Ωp+1 = αp+1 + (p + 1)αp ∧ β ∧ π. A sequential application of Theorem 1.5 leads to a constructive proof of the algebraic normal form of a two form which is useful for many arguments in the theory of exterior differential systems. Theorem 1.7. Let Ω ∈ Λ2 (V ∗ ) and let r be the smallest integer such that Ωr+1 = 0. Then there exist 2r linearly independent elements ω1 , . . . , ω2r such that (12)

Ω=

r

ωr+i ∧ ωi .

i=1

Proof. The theorem is proved by repeated applications of Theorem 1.5. In fact, from the hypotheses it follows that Ωr is decomposable and hence has a linear divisor ω1 . Next consider the ideal l(1) = {ω1 , Ω} generated by ω1 and Ω. Let r1 be the smallest integer such that Ωr1 +1 ∧ ω1 = 0.

§2. The Notion of an Exterior Differential System

13

Clearly r1 + 1 ≤ r. Then Ωr1 ∧ω1 is the Grassmann coordinate vector of the retraction space C(l(1)) and is decomposable and non-zero. Let ω2 be a linear factor of Ωr1 ∧ ω1 , which is linearly independent from ω1 . Then Ωr1 ∧ ω1 ∧ ω2 = 0. Let r2 be the smallest integer satisfying Ωr2 +1 ∧ ω1 ∧ ω2 = 0, so that r2 < r1 . Continuing this process, we get a sequence of positive integers r > r1 > r2 > . . . , which must end with zero. This means that there are linear forms ω1 , . . . , ωq , linearly independent, satisfying Ω ∧ ω1 ∧ · · · ∧ ωq = 0. From this we get Ω=

ηi ∧ ω i ,

1≤i≤q

where ηi are linear forms. Since Ω = 0, we must have q = r and ηi , ωi , 1 ≤ i ≤ r are linearly independent. The theorem is proved by setting r

ωr+i = ηi . Remark. Theorem 1.7 is equivalent to the theorem in linear algebra on the normal form of a skew-symmetric matrix. In fact, in terms of a base ωi , 1 ≤ i ≤ n, of V ∗ we can write Ω = 1/2 aij ωi ∧ ωj , aij + aji = 0. i,j

Let ωi =

sik ω∗ , k

1 ≤ i, k ≤ n

k

be a change of base. Then Ω = 1/2 where (13)

a∗kl =

aij sik sjl ,

a∗kl ω∗ ∧ ω∗ , k

l

1 ≤ i, j, k, l ≤ n.

i,j

If we introduce the matrices A = (aij ),

A∗ = (a∗ij ),

S = (sji ),

of which A and A∗ are skew-symmetric, and S is non-singular, then (13) can be written as a matrix equation A∗ = SAt S,

t

S = transpose of S.

Theorem 1.7 can be stated as follows: Given a skew-symmetric matrix A. Its rank is even. There exists a non-singular matrix S, such that ⎛ ⎞ 0 Ip 0 A∗ = ⎝ −Ip 0 0 ⎠ , 0 0 0 where Ip is the unit matrix of order p.

14

I. Preliminaries

§2. The Notion of an Exterior Differential System. Consider a differentiable manifold M of dimension n. Its cotangent bundle, whose fibers are the cotangent spaces Tx∗ (M ), x ∈ M , we will denote by T ∗ M . From T ∗ M we construct the bundle ΛT ∗ M , whose fibers are ΛTx∗ =

Λp Tx∗ ,

0≤p≤n

which have the structure of a graded algebra, as discussed in the last section. The bundle ΛT ∗ M has the subbundles Λp T ∗ M , whose definition is obvious. Similar definitions are valid for the tangent bundle T M . A section of the bundle Λp T ∗ M =

Λp Tx∗ → M

x∈M

is called an exterior differential form of degree p, or a form of degree p or simply a p-form. By abuse of language we will call a differential form a section of the bundle ΛT ∗ M ; its p-th component is a p-form. All sections are supposed to be sufficiently smooth. In terms of a system of local coordinates x1 , . . . , xn on M , an exterior differential form of degree p has the expression α = 1/p!

ai1 ...ip dxi1 ∧ · · · ∧ dxip ,

1 ≤ i1 , . . . , ip ≤ n,

where the coefficients are smooth functions and are anti-symmetric in any two of the indices. p Let Ωp (M ) = C ∞ -sections of Λp T ∗ M and let Ω∗ (M ) = Ω (M ). Definition. (i) An exterior differential system is given by an ideal I ⊂ Ω∗ (M ) that is closed under exterior differentiation; (ii) an integral manifold of the system is given by an immersion f : N → M such that f ∗ α = 0 for all α ∈ I. q By our conventions I = I is a direct sum of its homogeneous pieces I q = q I ∩ Ω (M ), and by differentiation; and by differential closure we have dα ∈ I whenever α ∈ I. We sometimes refer to an ideal I ⊂ Ω∗ (M ) satisfying dI ⊆ I as a differential ideal. In practice, I will be almost always generated as a differential ideal by a finite collection {αA}, 1 ≤ A ≤ N of differential forms; forms of degree zero, i.e. functions, are not excluded. An integral manifold of I is given by an immersion f :N →M satisfying f ∗ α = 0 for 1 ≤ A ≤ N . Then f ∗ (β ∧ αA) = 0 and f ∗ (dαA ) = 0, and so f ∗ α = 0 for all α in the differential ideal generated by the {αA}.

§2. The Notion of an Exterior Differential System

15

The fundamental problem in exterior differential systems is to study the integral manifolds. We may think of these as solutions to the system αA = 0 of exterior equations. When written out in local coordinates, this is a system of P.D.E.’s. The notion is of such generality that it includes all the ordinary and partial differential equations, as the following examples show: Example. The second-order differential equations in the (x, y)-plane, dy d2 y = F (x, y, ) dx2 dx can be written as an exterior differential system dy − y dx = 0, dy − F (x, y, y )dx = 0 in the space of the variables (x, y, y ). Example. Consider the partial differential equation of the first order F (xi , z,

(14)

∂z ) = 0, ∂xi

1 ≤ i ≤ n.

By introducing the partial derivatives as new variables, it can be written as an exterior differential system F (xi , z, pi ) = 0, dz − pi dxi = 0

(15)

in the (2n + 1)-dimensional space (xi , z, pi ). From these examples it is clear that any system of differential equations can be written as an exterior differential system. However, not all exterior differential systems arise in this way. The following example marks the birth of differential systems: Example. The equation a1 (x)dx1 + · · · + an (x)dxn = 0, x = (x1 , . . . , xn ), is called a Pfaffian equation. Pfaff’s problem is to determine its integral manifolds of maximal dimension. From the examples we notice two important concepts. One is an exterior differential system with independence condition (I, Ω) which is given by a closed differential ideal I together with a decomposable p-form Ω = ω1 ∧ · · · ∧ ωp .

16

I. Preliminaries

An integral manifold of (I, Ω) is an integral manifold of I satisfying the additional condition f ∗ Ω = 0. This is the case when we wish to keep some variables independent, as in the case when the system arises from a system of partial differential equations. For instance, in the second example, we take (16)

Ω = dx1 ∧ · · · ∧ dxn .

The partial differential equation (14) is equivalent to the system with independence condition (I, Ω), where I is generated by the left-hand members of (15) and Ω is given by (16). Whether an independence condition should be imposed depends on the particular problem. The other important concept is that of prolongation, which will be treated in detail later on. In our first and second examples it is necessary to introduce the derivatives as new variables. With more general systems the consideration of higherorder derivatives becomes necessary. Thus we could be forced to introduce higherdimensional manifolds and related systems, the prolonged systems, whose study is necessary for that of the given system. §3. Jet Bundles. A rigorous theory of differential systems depends on a foundation of differentiable manifolds and their differentiable maps. One such foundation is provided by the theory of jets developed by Charles Ehresmann. An introduction will be given below. We will give a geometric description of the spaces of partial derivatives of maps between two differentiable manifolds. These spaces will be constructed as differentiable manifolds with underlying sets given by equivalence classes of maps. The equivalence relation will be given in a form which is clearly intrinsic by first defining it for normalized functions on the real line and then defining it for general maps by a universal extension. The analytical content of the equivalence relations is then exhibited by a local characterization which is in turn used to provide the differentiable structure. Historically these ideas were motivated by geometers studying partial differential equations, say F (x1 , . . . , xm , z, ∂z/∂x1 , . . . , ∂z/∂xm ) = 0, and their desire to interpret this equation as representing a hypersurface in the space with coordinates x1 , . . . , xm , z, ∂z/∂x1 , . . . , ∂z/∂xm . The idea behind jets is simply to give this a precise formulation. Let R denote the real line, with the usual differentiable structure and let t denote a coordinate function in a neighborhood of the origin. If f : R → R and g : R → R are two differentiable maps of the real line into itself which map the origin into the origin, then f and g are said to have the same r-jet whenever dr g df dg dr f (0) = (0), . . . , r (0) = r (0). dt dt dt dt

§3. Jet Bundles

17

Now let N be a differentiable manifold, and let p ∈ N , then a p-based parametrized curve u, written u : (R, 0) → (N, p) is a map of the real line into N which takes the origin of R into p and is differentiable. Similarly, a p-based real valued function v, written v : (N, p) → (R, 0), is a real valued function on N which maps the point p onto the origin of R and is differentiable. Let M and N be differentiable manifolds and let f : N → M and g : N → M be differentiable maps of N into M . Then f and g are said to have the same r-jet at a point p ∈ M whenever a) f(p) = g(p) = q and b) for all p-based parametrized curves u : (R, 0) → (N, p), and for all q-based real valued functions v : (M, q) → (R, 0), the differentiable maps v◦f ◦u

and v ◦ g ◦ u

of the real line into itself mapping the origin into the origin have the same r-jet. The relation that two maps have the same r-jet at a point p is an equivalence relation and the equivalence class with the representative f :N →M will be denoted by jpr (f). The point p is called the source of jpr (f) and the point f(p) is called the target of jpr (f). We have given an intrinsic characterization of these equivalence classes. In order to get hold of this notion we express the relation in local coordinates. Given α = (α1 , . . . , αm), we define α! = α1 ! . . . αm ! and |α| = α1 + · · · + αm and given x = (x1 , . . . , xm ) we define xα = x1 and

α1

αm

. . . xm

∂ α1 ∂ αm . . . α α ∂x1 1 ∂xm m with the convention that Dx0 f = f(0). Dxα =

18

I. Preliminaries

Proposition 3.1. Let f and g be two differentiable maps f : R m → Rn

and

g : Rm → Rn

mapping the origin into the origin. Let {x1 , . . . , xm } denote coordinates in a neighborhood of the origin of Rm and {z 1 , . . . , z n } denote coordinates in a neighborhood of the origin of Rn . These coordinates allow us to introduce real valued functions f i , gi by f(x) = (f 1 (x), . . . , f n (x))

and

g(x) = (g1 (x), . . . , gn (x)).

With these notations f and g have the same r-jet at the origin if and only if Dxα f i (0) = Dxα gi (0)

(17)

(1 ≤ i ≤ n, |α| ≤ r).

Proof. Assume that (17) holds and let u : (R, 0) → (Rm , 0) be an arbitrary 0-based curve. Using the {x1 , . . . , xm } coordinates we may define u(t) = (u1 (t), . . . , um (t)). Next let v : (Rn , 0) → (R, 0) be an arbitrary 0-based real valued function. Then repeated application of the chain rule and the Leibniz product formula gives rise to an equation dk dγ u β α v ◦ f ◦ u| = F (D v(0), D f(0), (0)), t=0 z x dtk dtγ

|α|, |β|, γ ≤ r,

where F is a constant coefficient polynomial in the indicated indeterminates. It follows that dk dγ u β α v ◦ f ◦ u| = F (D v(0), D f(0), (0)) t=0 z x dtk dtγ dγ u = F (Dzβ v(0), Dxα g(0), γ (0)) dt =

dk v ◦ g ◦ u|t=0 dtk

for k ≤ r, which verifies that f and g have the same r-jet at 0. Conversely let us assume that j0r (f) = j0r (g). Then if we take u : (R, 0) → (Rm , 0)

§3. Jet Bundles

19

to be the 0-based parametrized curve defined by u(t) = (ξ1 t, . . . , ξm t) with ξi ∈ R, 1 ≤ i ≤ m and take vj : (Rn , 0) → (R, 0) to be the 0-based real valued function defined by projection on the j-th coordinate, i.e. vj (z 1 , . . . , z n ) = z j , 1 ≤ j ≤ n then by hypothesis dk j dk f (ξ1 t, . . . , ξm t) = k gj (ξ1 t, . . . , ξm t), k dt dt

k ≤ r,

which implies ∂x

1i1

∂k f j ∂ k gj i1 im im (0)ξ . . . ξ = (0)ξ1i1 . . . ξm , i m i 1 m 1 . . . ∂xm ∂x1 . . . ∂xmim i1 + · · · + im = k.

Since this last equation holds for all real ξ1 , . . . , ξm , the corresponding coefficients must be equal, that is Dxβ f j (0) = Dxβ gj (0),

1 ≤ j ≤ n,

|β| ≤ r

as claimed. In order to carry this last result over to the general situation f (N, p) −→ (M, q) g we introduce coordinates hU with p the origin, and hV with q the origin and define f (Rm , 0) −→ (Rn , 0) g by

f = hV ◦ f ◦ h−1 U

and

g = hV ◦ g ◦ h−1 U .

Clearly we have jpr (f) = jpr (g) if and only if j0r (f ) = j0r (g). r (N, M ) denote the set of all r-jets of mappings from N into M with Now let Jp,q source p and target q. Then define the set r J r (N, M ) = Jp,q (N, M ). p∈M,q∈N

20

I. Preliminaries

We introduce the natural projections α : J r (N, M ) → N

and β : J r (N, M ) → M

defined by α(jpr (f)) = p and β(jpr (f)) = f(p). Matters being so, if {U λ } denotes a coordinate covering of N and {V µ } denotes a coordinate covering for M , then we define a topology on the set J r (N, M ) by prescribing a coordinate covering to have underlying open sets W λµ = {jpr (f) | α(jpr (f)) ∈ U λ

and β(jpr (f)) ∈ V µ }.

Now if {x1 , . . . , xm } denote the coordinate functions on U λ and {z 1 , . . . , z n } denote the coordinate functions on V µ , then Proposition 3.1 implies that we may define a coordinate system on W λµ by (18) h(jpr (f)) = (xi (p), z j (f(p)), Dxα (z ◦ f)(p)), 1 ≤ i ≤ m, 1 ≤ j ≤ n, 1 ≤ |α| ≤ r. We will call these coordinates the natural coordinates on the jet space. The Leibniz product formula together with the chain rule guarantee that a differentiable change of local coordinates in U λ and in V µ will induce a differentiable change of local coordinates in J r (N, M ). The fact that this change of local coordinates has non-zero Jacobian determinant follows from the fact that the matrix is block upper triangular with the diagonal blocks given by symmetric powers of the Jacobian of the original coordinate change. Thus we have defined a differentiable structure on J r (N, M ). The next natural question is to determine the dimension of J r (N, M ) in terms of the dimensions of N and M . A real valued function on an m-dimensional manifold N f :N →R has as many derivatives of order i as there are independent homogeneous polynomials of degree i. This number is

m+i−1 m−1

=

m+i−1 i

.

The total dimension of J r (N, R) is thus given by m+1+

r m+i−1 i=1

i

=m+

r m+i−1 i=0

i

=m+

m+r r

.

Now each coordinate function in a target space M will give rise to an independent set of derivatives, thus dim J r (N, M ) = dim N + dim M

dim N + r r

.

§3. Jet Bundles

21

Example. J 2 (R2 , R) Let (x, y) denote coordinates on R2 and z a coordinate on R. Let 2 (f)) = ∂f/∂x p(j(x,y)

and

2 q(j(x,y) (f)) = ∂f/∂y

and 2 (f)) = ∂ 2 f/∂x2 , r(j(x,y)

2 s(j(x,y) (f)) = ∂ 2 f/∂x∂y,

2 t(j(x,y) (f)) = ∂ 2 f/∂y2 .

Then 2 h(j(x,y) (f)) = (x, y, z, p, q, r, s, t)

defines the natural coordinates for J 1 (R2 , R). If we introduce a change of coordinate on R2 by S(x, y) = (ξ(x, y), η(x, y)), then this induces a transformation of the derivatives. In fact ⎛ ⎞ ⎛ ⎞ r r ⎜ s ⎟ S 2 (J(S)) | H(S) ⎜ s ⎟ ⎜ ⎟ ⎜ ⎟ ⎜t⎟= ⎜ t ⎟, ⎝ ⎠ 0 | J(S) ⎝ ⎠ p p q q where J(S) =

∂ξ/∂x ∂ξ/∂y

∂η/∂x ∂η/∂y

⎛

∂ 2 ξ/∂x2 ⎝ H(S) = ∂ 2 ξ/∂x∂y ∂ 2 ξ/∂y2

⎞ ∂ 2 η/∂x2 ∂ 2 η/∂x∂y ⎠ ∂ 2 η/∂y2

and ⎛

(∂ξ/∂x)2 2 S (J(S))(S) = ⎝ ∂ξ/∂x ∂ξ/∂y (∂ξ/∂y)2

2∂ξ/∂x ∂η/∂x ∂ξ/∂x ∂η/∂y + ∂ξ/∂y ∂η/∂x 2∂ξ/∂y ∂η/∂y

⎞ (∂η/∂x2 ) ∂η/∂x ∂η/∂y ⎠ . (∂η/∂y)2

A good viewpoint to keep in mind is that J r (N, M ) → N × M, that is, J r (N, M ) sits over N × M , and the coordinate transformations on N × M induce the action of a linear group on the set of elements in the inverse image of a point. The notion of jet bundles allows us to formulate general problems in differential geometry. As an example we observe that a partial differential equation for maps f :N →M can be described by an imbedded submanifold i : Σ → J r (N, M ).

22

I. Preliminaries

A solution is a map f : N → M such that for all p ∈ N jpr (f) ∈ i(Σ). We introduce the r-graph of a map f j r (f) : N → J r (N, M ) by the definition j r (f)(p) = jpr (f). Then the problem of finding solutions to a partial differential equation is the problem of finding maps whose r-graphs lie on the locus i(Σ) of the partial differential equation. This will be illustrated in Chapters IX and X. Several standard constructions of differential geometry fit into the language of jet bundles. For example the cotangent space Tp∗ (N ) for p ∈ N is defined by 1 (N, R) Tp∗ (N ) = Jp,0

and the differential of a real valued function f : N → R at p ∈ N is defined by df|p = jp1 (f − f(p)). The vector space structure on Tp∗ (N ) is intrinsically induced from the real line by αjp1 (f) + βjp1 (g) = jp1 (αf + βg). The tangent space Tp (N ) for p ∈ N is defined as the space of linear functionals on Tp∗ (N ) and is realized by 1 Tp (N ) = J0,p (R, N ) under the action j01 (u), jp1 (f) = d/dt(f ◦ u)|t=0 . In particular the cotangent bundle is defined by T ∗ (N ) = Tp∗ (N ) ⊂ J 1 (N, R) p∈M

and the tangent bundle is defined by T (N ) = Tp (N ) ⊂ J 1 (R, N ). p∈m

Finally we wish to introduce the contact system Ωr (N, M ) of a jet bundle J (N, M ). By a change of notation we can write the natural coordinates in (18) as r

(19)

α α xi (p), z α (f(p)), pα i , pi1i2 , . . . , pi1...ir

1 ≤ i, i1 , . . . , ir ≤ m,

1 ≤ α ≤ n,

where the p’s are the partial derivatives with respect to the xi ’s, up to the order r inclusive, and are symmetric in their lower indices. The Pfaffian equations i dz α − pα i dx = 0, i2 pα dpα (20) i1 − i i dx = 0, α1 2 α dpi1 ...ir−1 − pi1 ...ir−1 ir dxir = 0, define the contact system Ωr (N, M ). A form in Ωr (N, M ) is called a contact form. The forms (20) are those that naturally arise when a system of partial differential equations is converted to an exterior differential system. The fundamental property of these systems is contained in the following theorem.

§3. Jet Bundles

23

Theorem 3.2. A section σ : N → J r (N, M ) is a r-graph, that is σ(p) = jpr (f), if and only if σ ∗ Ωr (N, M ) = 0. A proof of this theorem and an intrinsic treatment of the contact system can be found in various sources, cf. Gardner and Shadwick [1987] or Goldschmidt and Sternberg [1973].

24

II. Basic Theorems

CHAPTER II BASIC THEOREMS

In this chapter we consider classical results on simple exterior differential systems that can be established by algebra and ordinary differential equations. In particular these results hold in the C ∞ -category. This is to be contrasted with later results that rely on the Cartan–K¨ ahler theorem and hold in the analytic category. §1. Frobenius Theorem. Perhaps the simplest exterior differential systems are those whose differential ideal I is generated algebraically by forms of degree one. Let the generators be α1 , . . . , αn−r , which we suppose to be linearly independent. The condition that I is closed gives (F )

dαi ≡ 0,

mod α1 , . . . , αn−r ,

1 ≤ i ≤ n − r.

This condition (F ) is called the Frobenius condition. A differential system α1 = · · · = αn−r = 0 satisfying (F ) is called completely integrable. Geometrically the α’s span at every point x ∈ M a subspace Wx of dimension n − r in the cotangent space Tx∗ (M ) or, what is the same, a subspace Wx⊥ of dimension r in the tangent space Tx . Following Chevalley, such data is known as a distribution. Notice that the condition (F ) is intrinsic, i.e., independent of local coordinates, and is also invariant under a linear change of the α’s with C ∞ coefficients. The fundamental theorem on completely integrable systems is: Theorem 1.1 (Frobenius). Let I be a differential ideal having as generators the linearly independent forms α1 , . . . , αn−r of degree one, so that the condition (F ) is satisfied. In a sufficiently small neighborhood there is a coordinate system y1 , . . . , yn such that I is generated by dyr+1 , . . . , dyn . Proof. We will prove the theorem by induction on r. Let r = 1. Then the subspace Wx⊥ ⊂ Tx , x ∈ M , is of dimension 1. Relative to a system of local coordinates xi , 1 ≤ i ≤ n, the equations of the differential system is written in the classical form dx1 dxn = · · · = , X 1 (x) X n (x) where the functions X i (x1 , . . . , xn ), not all zero, are the coefficients of a vector field X = i X i (x)∂/∂xi spanning Wx⊥ . By the flow box coordinate theorem (Warner

§1. Frobenius Theorem

25

[1971], p. 40), we can choose coordinates y1 , . . . , yn , such that Wx⊥ is spanned by the vector ∂/∂y1 ; then Wx is spanned by dy2 , . . . , dyn . The latter clearly form a set of generators of I. Notice that in this case the condition (F ) is void. Suppose r ≥ 2 and the theorem be true for r − 1. Let xi , 1 ≤ i ≤ n, be local coordinates such that α1 , . . . , αn−r , dxr are linearly independent. The differential system defined by these n − r + 1 forms also satisfies the condition (F ). By the induction hypothesis there are coordinates yi so that dyr , dyr+1 , . . . , dyn are a set of generators of the corresponding differential ideal. It follows that dxr is a linear combination of these forms or that xr is a function of yr , . . . , yn . Without loss of generality we suppose ∂xr /∂yr = 0. Since dxr =

∂xr r ∂xr dy + dyr+i , ∂yr ∂yr+1

1≤ i ≤ n−r

i

we may now solve for dyr in terms of dxr and dyr+1 , . . . , dyn . Since α1 , . . . , αn−r are linear combinations of dyr , . . . , dyn they can now be expressed in the form αi =

aij dyr+j + bi dxr ,

1 ≤ i, j ≤ n − r.

j

Since αi and dxr are linearly independent, the matrix (aij ) must be non-singular. Hence we can find a new set of generators for I in the form αi = dyr+i + pi dxr ,

1 ≤ i ≤ n − r,

and the condition (F ) remains satisfied. Exterior differentiation gives dαi = dpi ∧ dxr ≡

1≤λ≤r−1

∂pi λ dy ∧ dxr ≡ 0, ∂yλ

mod α1 , . . . , αn−r .

It follows that ∂pi /∂yλ = 0,

1 ≤ i ≤ n − r,

1 ≤ λ ≤ r − 1,

which means that pi are functions of yr , . . . , yn . Hence in the y-coordinates we are studying a system of n − r forms of degree one involving only the n − r + 1 coordinates yr , . . . , yn . This reduces to the situation settled at the beginning of this proof. Hence the induction is complete. The theorem gives a “normal form” of a completely integrable system, i.e., the system can be written locally as dyr+1 = · · · = dyn = 0

26

II. Basic Theorems

in a suitable coordinate system. The maximal integral manifolds are yr+1 = const, . . . , yn = const, and are therefore of dimension r. We say that the system defines a foliation, of dimension r and codimension n − r, of which these submanifolds are the leaves. The simplest non-trivial case of the Frobenius theorem is the system generated by a single one form in three space. Thus I = {Rdx + Sdy + T dz} and the condition (F ) are the necessary and sufficient conditions that there exist an integrating factor for the one form ω = Rdx + Sdy + T dz. That is there exists a function µ such that µω is exact. This example will be considered when condition (F ) is not identically satisfied in Example 5.11 of Chapter IV. The condition (F ) has a formulation in terms of vector fields, which is also useful. We add to α1 , . . . , αn−r the r forms αn−r+1 , . . . , αn , so that αi , 1 ≤ i ≤ n, are linearly independent. Then we have (1) dαi = 1/2 cijk αj ∧ αk , 1 ≤ i, j, k ≤ n, cijk + cikj = 0. j,k

The condition (F ) can be expressed as (2)

capq = 0,

1 ≤ a ≤ n − r,

n − r + 1 ≤ p, q ≤ n.

Let f be a smooth function. The equation (3) df = (Xi f)αi defines n operators or vector fields Xi , which form a dual base to αi . Exterior differentiation of (3) gives 1/2 (Xi (Xj (f)) − Xj (Xi (f)))αi ∧ αj + Xi (f)dαi = 0. i,j

j

Substituting (1) into this equation, we get (4)

[Xi , Xj ]f = (Xi Xj − Xj Xi )f = −

ckij Xk f.

It follows that the condition (2) can be written (5) [Xp , Xq ]f = − cspq Xs f, n − r + 1 ≤ p, q, s ≤ n. Equation (4) is the dual version of (1). The vectors Xn−r+1 , . . . , Xn span at each point x ∈ M the subspace Wx⊥ of the distribution. Hence the condition (F ) or (2) or (5) can be expressed as follows: Proposition 1.2. Let a distribution M be defined by the subspace Wx⊥ ⊂ Tx , dim Wx⊥ = r. The condition (F ) says that, for any two vector fields X, Y , such that Xx , Yx ∈ Wx⊥ , their bracket [X, Y ]x ∈ Wx⊥ .

§2. Cauchy Characteristics

27

§2. Cauchy Characteristics. The Frobenius Theorem shows that a completely integrable system takes a very simple form upon a proper choice of the local coordinates. Given any exterior differential system, one can ask the question whether there is a coordinate system such that the system is generated by forms in a smaller number of these coordinates. This question is answered by the Cauchy characteristics. Its algebraic basis is the retraction theorem (Theorem 1.3 of Chapter I). Let I be a differential ideal. A vector field ξ such that ξ I ⊂ I is called a Cauchy characteristic vector field of I. At a point x ∈ M we define A(I)x = {ξx ∈ Tx M | ξx

Ix ⊂ Ix }

∗ and C(I)x = A(I)⊥ x ⊂ Tx M . These concepts reduce to the ones treated in §1, Chapter I. In particular, we will call C(I)x the retracting space at x and call dim C(I)x the class of I at x. We have now a family of ideals Ix depending on the parameter x ∈ M . When restricting to a point x we have a purely algebraic situation.

Proposition 2.1. If ξ, η are Cauchy characteristic vector fields of a differential ideal I, so is their bracket [ξ, η]. Proof. Let Lξ be the Lie derivative defined by ξ. It is well-known Lξ = d(ξ ) + (ξ )d. Since I is closed, we have dI ⊂ I. If ξ is a characteristic vector field, we have ξ I ⊂ I. It follows that Lξ I ⊂ I. The lemma follows from the identity (6)

[Lξ , η ] = Lξ η def

−η

Lξ = [ξ, η] ,

which is valid for any two vector fields ξ, η. To prove (6) we observe that Lξ is a derivation of degree 0 and η is a derivation of degree −1, so that [Lξ , η ] is also a derivation of degree −1. It therefore suffices to verify (6) when the two sides act on functions f and differentials df. Clearly, when acting on f, both sides give zero. When acting on df, we have [Lξ , η ]df = Lξ (ηf) − η

d(ξf)

= [ξ, η]f = [ξ, η]

df.

This proves (6) and hence the proposition.

Theorem 2.2. Let I be a finitely generated differential ideal whose retracting space C(I) has constant dimension s = n − r. Then there is a neighborhood in which there are coordinates (x1 , . . . , xr ; y1 , . . . , ys ) such that I has a set of generators that are forms in y1 , . . . , ys and their differentials. Proof. By Proposition 1.2 the differential system defined by C(I) (or what is the same, the distribution defined by A(I)) is completely integrable. We may choose coordinates (x1 , . . . , xr ; y1 , . . . , ys ) so that the foliation so defined is given by yσ = const,

1 ≤ σ ≤ s.

28

II. Basic Theorems

By the retraction theorem, I has a set of generators which are forms in dyσ , 1 ≤ σ ≤ s. But their coefficients may involve xρ , 1 ≤ ρ ≤ r. The theorem follows when we show that we can choose a new set of generators for I which are forms in the yσ coordinates in which the xρ do not enter. To exclude the trivial case we suppose the I is a proper ideal, so that it contains no non-zero functions. Let Iq be the set of q-forms in I, q = 1, 2, . . . . Let ϕ1 , . . . , ϕp be linearly independent 1-forms in I1 such that any form in I1 is their linear combination. ∂ Since I is closed, dϕi ∈ I, 1 ≤ i ≤ p. For a fixed ρ we have ∂x ρ ∈ A(I), which implies ∂ dϕi = L∂/∂xρ ϕi ∈ I1 , ∂xρ since the left-hand side is of degree 1. It follows that (7)

∂ϕi i ρϕ = = L aij ϕj , ∂/∂x ∂xρ

1 ≤ i, j ≤ p

j

where the left-hand side stands for the form obtained from ϕi by taking the partial derivatives of the coefficients with respect to xρ . For this fixed ρ we regard xρ as the variable and 1 ρ−1 ρ+1 r 1 s x , . . . , x , x , . . . , x , y , . . . , y as parameters. Consider the system of ordinary differential equations (8)

dz i = aij z j , dxρ

1 ≤ i, j ≤ p.

j

i Let z(k) , 1 ≤ k ≤ p, be a fundamental system of solutions, so that i det(z(k) ) = 0.

We shall replace ϕi by the ϕ˜k defined by (9)

ϕi =

i z(k) ϕ˜k .

By differentiating (9) with respect to xρ and using (7), (8), we get ∂ ϕ˜k = 0, ∂xρ so that ϕ ˜k does not involve xρ . Applying the same process to the other x’s, we arrive at a set of generators of I1 which are forms in yσ . Suppose this process carried out for I1 , . . . , Iq−1, so that they consist of forms in yσ . Let Jq−1 be the ideal generated by I1 , . . . , Iq−1 . Let ψα ∈ Iq , 1 ≤ α ≤ r, be linearly independent mod Jq−1 , such that any q-form of Iq is congruent mod Jq−1 to a linear combination of them. By the above argument such forms include ∂ dψα = L∂/∂xρ ψα . ∂xρ


29

Hence we have ∂ψα β ≡ bα βψ , ρ ∂x

mod Jq−1 ,

1 ≤ α, β ≤ r.

By using the above argument, we can replace the ψα by ψ˜β such that ∂ ψ˜α ∈ Jq−1 . ∂xρ This means that we can write ∂ ψ˜α = ηhα ∧ ωhα , ρ ∂x h

where ηhα ∈ I1 ∪ · · · ∪ Iq−1 and are therefore forms in yσ . Let θhα be defined by ∂θhα = ωhα . ∂xρ Then the forms

≈

ψα = ψ˜α −

ηhα ∧ θhα

h ρ

do not involve x , and can be used to replace ψα . Applying this process to all xρ , 1 ≤ ρ ≤ r, we find a set of generators for Iq , which are forms in yσ only. Definition. The leaves defined by the distribution A(I) are called the Cauchy characteristics. Notice that generally r is zero, so that a differential system generally does not have Cauchy characteristics (i.e., they are points). The above theorem allows us to locally reduce a differential ideal to a system in which there are no extraneous variables in the sense that all coordinates are needed to express I in any coordinate system. Thus the class of I equals the minimal number of variables needed to describe the system. An often useful corollary of Theorem 2.2 which illustrates its geometric content is the following: Corollary 2.3. Let f : M → M be a submersion with vertical distribution V ⊂ T (M ) with connected fibers over x ∈ M given by (ker f∗ )x . Then a form α on M is the pull-back f ∗ α of a form α on M if and only if v

α=0

and

v

dα = 0

for all

v ∈ V.

Proof. By the submersion theorem (Warner [1971], p. 31), there are local coordinates such that f(x1 , . . . , xp , xp+1 , . . . , xN ) = (x1 , . . . , xp ).

As such V =

∂ ∂ ,··· , N p+1 ∂x ∂x

.

30

II. Basic Theorems

Now setting I = (α), we see that V ⊂ A(I). Therefore, by Theorem 2.2 there exists a generator for I independent of (xp+1 , . . . , xN ), and hence of the form f ∗ α with α ∈ M . Thus there is a function µ such that µα = f ∗ α . Since 0=v

(dµ ∧ α + µdα ) = v(µ)α

for all v ∈ V,

we see that µ is independent of (xp+1 , . . . , xN ) and hence µ = λ◦f for some function λ defined on M . Setting α = λ1 α we have our result that α = f ∗ (α ). We will apply this theorem to the first order partial differential equation F (xi , z, ∂z/∂xi ) = 0,

1 ≤ i ≤ n.

Following the example starting with §2 equation (14) of Chapter I, the equation can be formulated as the differential system (15), §2, Chapter I. To these equations we add their exterior derivatives to obtain

(10)

F (xi , z, pi ) = 0 dz − pi dxi = 0 (Fxi + Fz pi )dxi + Fpi dpi = 0 dxi ∧ dpi = 0.

These equations are in the (2n+1)-dimensional space (xi , z, pi ). The corresponding differential ideal is generated by the left-hand members of (10). To determine the space A(I) consider the vector ξ=

ui ∂/∂xi + u∂/∂z +

vi ∂/∂pi

and express the condition that the interior product ξ This gives

(11)

keeps the ideal I stable.

u− pi ui = 0, (Fxi + Fz pi )ui + Fpi vi = 0, (ui dpi − vi dxi ) = 0.

Comparing the last equation of (11) with the third equation (10) we get (12)

ui = λFpi ,

vi = −λ(Fxi + Fz pi ),

and the first equation of (11) then gives (13)

u=λ

pi Fpi .


31

The parameter λ being arbitrary, equations (12) and (13) show that dim A(I) = 1, i.e., the characteristic vectors at each point form a one-dimensional space. This fundamental (and remarkable) fact is the key to the theory of partial differential equations of the first order. The characteristic curves in the space (xi , z, pi), or characteristic strips in the classical terminology, are the integral curves of the differential system (14)

dxi dpi dz =− = . Fpi F x i + F z pi pi Fpi

These are the equations of Charpit and Lagrange. To construct an integral manifold of dimension n it suffices to take an (n−1)-dimensional integral manifold transverse to the Cauchy characteristic vector field (or non-characteristic data in the classical terminology) and draw the characteristic strips through its points. Putting it in another way, an n-dimensional integral manifold is generated by characteristic strips. We remark that points in (xi , pi )-space may be thought of as hyperplanes pi dxi = 0 in the tangent spaces Tx (Rn ). A curve in (xi , z, pi )-space projects to a curve in (xi , pi )-space, which is geometrically a 1-parameter family of tangent hyperplanes. This is the meaning of the terminology “strips”. Example. Consider the initial value problem for the partial differential equation ∂z ∂z + =1 ∂x ∂y √ with initial data given along y = 0 by z(x, 0) = x. Let us introduce natural coordinates√ in J 1 (2, 1) by (x, y, z, p, q). This initial data D : R → R2 × R where D(x) = (x, 0, x) is extended to a map δ : R → J 2 (2, 1) where the image satisfies the equation and the strip condition z

1 0 = δ ∗ (dz − p dx − q dy) = √ dx − p dx 2 x 1 and q = 1 − zp = 12 and δ is unique. (For the general non-linear here p = 2√ x equation, there can be more than one choice of δ.) The extended data becomes

√ 1 δ(x) = (x, 0, x, √ , 1/2). 2 x If we parametrize the equation by i : Σ → J 1 (2, 1) where i(x, y, z, p) = (x, y, z, p, 1− zp), then the data can be pulled back to a map ∆ : R → Σ, where ∆(s) = √ 1 (s, 0, s, 2√ ). s The Cauchy characteristic vector field is X=z

∂ ∂ ∂ ∂ + + − p2 ∂x ∂y ∂z ∂p

and the corresponding flow is given by dx = z, dt

dy = 1, dt

dz = 1, dt

dp = −p2 . dt

32

II. Basic Theorems

The solution for the given data representing the union of characteristic curves along the data is √ √ t2 x= + ( s)t + s, y = t, z = t + s, 2 and eliminating s and t gives an implicit equation for z(x, y), namely z 2 − zy = x −

y2 . 2

Note that only the upper branch of the double-valued solution

y ± 4x − y2 z= 2 actually satisfies the initial conditions. Next we wish to apply the Cauchy characteristics to prove the following global theorem: Theorem 2.4. Consider the eikonal differential equation (15) (∂z/∂xi )2 = 1 1 ≤ i ≤ n. If z = z(x1 , . . . , xn ) is a solution valid for all (x1 , . . . , xn) ∈ (= n-dimensional euclidean space), then z is a linear function in xi , i.e., z= ai xi + b, where ai , b are constants satisfying

En

a2i = 1.

Proof. We will denote by E n+1 the space of (x1 , . . . , xn , z), and identify E n with the hyperplane z = 0. The solution can be interpreted as a graph Γ in E n+1 having a one-one projection to E n . For the equation (15) the denominators in the middle term of (14) are zero, so that the Cauchy characteristics satisfy pi = const. The equations (14) can be integrated and the Cauchy characteristic curves, when projected to E n+1 , are the straight lines (16)

xi = xi0 + pi t,

z = z0 + t,

where xi0 , pi z0 are constants. Hence the graph Γ must have the property that it is generated by the “Cauchy lines” (16), whose projections in E n form a foliation of En. Changing the notation in the first equation of (16), we write it as x∗ = xi + i

∂z t, ∂xi

where z = z(x1 , . . . , xn ) is a solution of (15). For a given t ∈ R this can be interpreted as a diffeomorphism ft : E n → E n defined by ft (x) = x∗ = (x∗ , . . . , x∗ ), 1

n

x, x∗ ∈ E n .

§3. Theorems of Pfaff and Darboux

33

Geometrically it maps x ∈ E n to the point x∗ at a distance t along the Cauchy line through x; this makes sense, because the Cauchy lines are oriented. Its Jacobian determinant is ∂2z i J (t) = det δj + t ∂xi ∂xj and is never zero. But this implies (17)

∂2z = 0, ∂xi ∂xj

and hence that z is linear. For if (17) is not true, then the symmetric matrix (∂ 2 z/∂xi ∂xj ) has a real non-zero eigenvalue, say λ, and J (−1/λ) = 0, which is a contradiction. Remark. The function

z=

1/2 i 2

(x )

i

satisfies (15), except at xi = 0. Hence Theorem 2.3 needs the hypothesis that (15) is valid for all x ∈ E n .

§3. Theorems of Pfaff and Darboux. Another simple exterior differential system is one which consists of a single equation (18)

α = 0,

where α is a form of degree 1. This problem was studied by Pfaff [1814-15]. The corresponding closed differential ideal I has the generators α, dα. The integer r defined by (19)

(dα)r ∧ α = 0,

(dα)r+1 ∧ α = 0

is called the rank of the equation (18). It depends on the point x ∈ M , and is invariant under the change α → aα, a = 0. Putting it in a different way, the two-form dα mod α, has an even rank 2r in the sense of linear algebra. The study of the integral manifolds of (18) is clarified by the normal form, given by the Theorem 3.1 (The Pfaff problem). In a neighborhood suppose the equation (18) has constant rank r. Then there exists a coordinate system w 1 , . . . , wn , possibly in a smaller neighborhood, such that the equation becomes (20)

dw 1 + w 2 dw 3 + · · · + w 2r dw 2r+1 = 0.

34

II. Basic Theorems

Proof. Let I = {α, dα} be the ideal generated by α, dα. By Theorem 1.5 of Chapter I and (19), the retraction space C(I) is of dimension 2r + 1 and has the Grassmann coordinate vector (dα)r ∧ α. By Theorem 2.2 there is a function f1 such that (dα)r ∧ α ∧ df1 = 0. Next let I1 be the ideal {df1 , α, dα}. If r = 0, our theorem follows from the Frobenius theorem. If r > 0, the forms df1 and α must be linearly independent. Applying Theorem 1.5, Chapter I to I1 , let r1 be the smallest integer such that (dα)r1 +1 ∧ α ∧ df1 = 0. Clearly r1 +1 ≤ r. The equality sign must hold, as otherwise we get a contradiction to the first equation of (19), by Theorem 1.6, Chapter I. Applying Theorem 2.2 to I1 , there is a function f2 such that (dα)r−1 ∧ α ∧ df1 ∧ df2 = 0. Continuing this process, we find r functions f1 , . . . , fr satisfying dα ∧ α ∧ df1 ∧ · · · ∧ dfr = 0, α ∧ df1 ∧ · · · ∧ dfr = 0. Finally, let Ir be the ideal {df1 , . . . , dfr , α, dα}. Its retraction space C(Ir ) is of dimension r + 1. There is a function fr+1 such that α ∧ df1 ∧ · · · ∧ dfr+1 = 0, df1 ∧ · · · ∧ dfr+1 = 0. By modifying α by a factor, we can write α = dfr+1 + g1 df1 + · · · + gr dfr . Because of the first equation of (19) the functions f1 , . . . , fr+1 , g1 , . . . , gr are independent. Theorem 3.1 follows by setting w 1 = fr+1 ,

w 2i = gi ,

w 2i+1 = fi ,

1 ≤ i ≤ r.

Corollary 3.2 (Symmetric normal form). In a neighborhood suppose the equation (18) has constant rank r. Then there exist independent functions z, y1 , . . . , yr , x1 , . . . , xr such that the equation becomes (21)

dz + 1/2

r

(yi dxi − xi dyi ) = 0.

i=1

Proof. It suffices to apply the change of coordinates w1 = z − w 2i = yi ,

w 2i+1

1 i i xy, 2 = xi , 1 ≤ i ≤ r.

§3. Theorems of Pfaff and Darboux

35

From the normal form (20) we see that the maximal integral manifolds are of dimension r. They are, for instance, given by w 1 = f(w 3 , . . . , w2s+1 ), w 2t+1 = const,

s 3. Even when (3) is not overdetermined, it cannot generally be placed in one of the classical forms (e.g., Cauchy–Kowalevski). Nevertheless, certain systems of equations of the form (3) had been treated successfully (at least, in the real analytic category) in the nineteenth century by methods generalizing the initial value problem (sometimes called the “Cauchy problem” because of Cauchy’s work on initial value problems). Let us illustrate such an approach by the following simple example: Consider the following system of first order partial differential equations for one function u of two variables x and y: (4)

ux = F (x, y, u),

uy = G(x, y, u).

If we seek a solution of (4) which satisfies u(0, 0) = c, then we may try to construct such a solution by first solving the initial value problem (5)

vx = F (x, 0, v) where

v(0) = c

for v as a function of x, and then solving the initial value problem (regarding x as a parameter) (6)

uy = G(x, y, u) where

u(x, 0) = v(x).

Assuming that F and G are smooth in a neighborhood of (x, y, u) = (0, 0, c), standard O.D.E. theory tells us that this process will yield a smooth function u(x, y) defined on a neighborhood of (x, y) = (0, 0). However, the function u may not satisfy the equation ux = F (x, y, u) except along the line y = 0. In fact, if we set E(x, y) = ux(x, y) − F (x, y, u), then E(x, 0) = 0 and we may compute that Ey (x, y) = (G(x, y, u))x − Fy (x, y, u) − Fu (x, y, u)G(x, y, u) = Gu (x, y, u)E(x, y) + T (x, y, u)

Introduction

53

where T (x, y, u) = F (x, y, u)Gu(x, y, u) − G(x, y, u)Fu (x, y, u) + Gx (x, y, u) − Fy (x, y, u). Suppose that F and G satisfy the identity T ≡ 0. Then E satisfies the differential equation with initial condition Ey = Gu (x, y, u)E

and E(x, 0) = 0.

By the usual uniqueness theorem in O.D.E., it follows that E(x, y) ≡ 0, so u satisfies the system of equations (4). It follows that the condition T ≡ 0 is a sufficient condition for the existence of local solutions of (4) where u(0, 0) is allowed to be an arbitrary constant as long as (0, 0, u(0, 0)) is in the common domain of F and G. Note that if we consider the differential system I (on the domain in R3 where F and G are both defined) which is generated by the 1-form ϑ = du − F dx − Gdy, then dϑ ≡ −T dx ∧ dy mod ϑ, so the condition T ≡ 0 is equivalent to the condition that I be generated algebraically by ϑ. Thus, we recover a special case of the Frobenius theorem. It is an important observation that the process of computing the differential closure of this system uncovers the “compatibility condition” T ≡ 0. Let us pursue the case of first order equations with two independent variables a little further. Given a system of P.D.E. R(x, y, u, ux, uy ) = 0, where u is regarded as a vector-valued function of the independent variables x and y, then, under certain mild constant rank assumptions, it will be possible to place the equations in the following (local) normal form (i)

u0x = F (x, y, u)

(ii)

u0y = G(x, y, u, ux) u1y = H(x, y, u, ux)

by making suitable changes of coordinates in (x, y) and decomposing u into u = (u0 , u1, u2 ) where each of the uα is a (vector-valued) unknown function of x and y. Note that the original system may thus be (roughly) regarded as being composed of an “overdetermined” part (for u0 ), a “determined” part (for u1 ), and an “underdetermined” part (for u2 ). (This “normal form” generalizes in a straightforward way to the case of n independent variables, in which case the unknown functions u are split into (n + 1) vector-valued components.) The “Cauchy–Kowalewski approach” to solving this system in the real analytic case can then be described as follows: Suppose that the collection uα consists of sα ≥ 0 unknown functions. For simplicity’s sake, we assume that F , G, H are real analytic and well-defined on the entire Rk (where k has the appropriate dimension). Then we choose s0 constants, which we write as f 0 , s1 analytic functions of x, which we write as f 1 (x), and s2 analytic functions of x and y, which we write as f 2 (x, y). We then first solve the following system of O.D.E. with initial conditions for s0 functions v0 of x: (i )

vx0 = F (x, 0, v0, f 1 (x), f 2 (x, 0)) v0 (0) = f 0

54

III. Cartan–K¨ ahler Theory

and then second solve the following system of P.D.E. with initial conditions for s0 + s1 functions (u0 , u1 ) of x and y: u0y = G(x, y, u0 , u1, f 2 (x, y), u0x , u1x, fx2 (x, y)) (ii )

u1y = H(x, y, u0 , u1 , f 2 (x, y), u0x , u1x, fx2 (x, y)) u0 (x, 0) = v0 (x) u1 (x, 0) = f 1 (x).

This process yields a function u(x, y) = (u0 (x, y), u1 (x, y), u2 (x, y)) (where u2 (x, y) is defined to be f 2 (x, y)) which is uniquely determined by the collection f = {f 0 , f 1 (x), f 2 (x, y)}. While it is clear that the u(x, y) thus constructed satisfies (ii), it is not at all clear that u satisfies (i). In fact, if we set E(x, y) = u0x (x, y) − f(x, y, u(x, y)), then E(x, 0) ≡ 0 since u(x, 0) satisfies (i ), but, in general E(x, y) ≡ 0 for the generic choice of “initial data” f. In the classical terminology, the system (i), (ii) is said to be “involutive” or “in involution” if, for arbitrary analytic initial data f, the unique solution uf of (i ,ii ) is also a solution of (i,ii). Because of the nature of the initial conditions f, the classical terminology further described the “general solution” of (i,ii) in the involutive case as “depending on s0 constants, s1 functions of one variable, and s2 functions of two variables”. In the analytic category, the condition of involutivity for the system (i,ii) can be expressed in terms of certain P.D.E., called “compatibility conditions”, which must be satisfied by the functions F , G, and H. For example, in the case of (4), the compatibility condition takes the form T ≡ 0. Also note that, for the system (4), we have (s0 , s1 , s2 ) = (1, 0, 0). Of course, this notion of involutivity extends to P.D.E. systems with n independent variables. The condition of involutivity is rather stringent (except in the case (s0 , . . . , sn ) = (0, . . . , 0, s, 0), which corresponds to the classical Cauchy problem). Thus, one often must modify the equations in some way in order to reduce to the involutive case. Let us give an example. Consider the following system of three equations for three unknown functions u1 , u2 , u3 of three independent variables x1 , x2, x3 . Here we write ∂j for ∂/∂xj and v1 , v2 , v3 are some given functions of x1 , x2 , x3 . ∂2 u3 − ∂3 u2 = u1 + v1 (7, i,ii,iii)

∂3 u1 − ∂1 u3 = u2 + v2 ∂1 u2 − ∂2 u1 = u3 + v3

The approach to treating (7) as a sequence of Cauchy problems (with (s0 , s1 , s2 , s3 ) = (0, 1, 1, 1)) is as follows: (1) Choose three functions ϕ1 (x1 ), ϕ2 (x1 , x2 ), ϕ3 (x1 , x2 , x3 ). (2) Solve the equation in R2 , ∂2 w = ∂1 ϕ2 − ϕ¯3 − v¯3 with the initial condition w(x1 , 0) = ϕ1 (x1 ) where ϕ¯3 (x1 , x2) = ϕ3 (x1 , x2 , 0) and v¯3 (x1 , x2) = v3 (x1 , x2, 0).

Introduction

55

(3) Solve the pair of equations ∂3 u1 = ∂1 ϕ3 + u2 + v2 and ∂3 u2 = ∂2 ϕ3 − u1 − v1 with the initial conditions u1 (x1 , x2 , 0) = w(x1 , x2 ) and u2 (x1 , x2, 0) = ϕ2 (x1 , x2 ). (4) Set u3 equal to ϕ3 . However, the resulting set of functions ua will not generally be a solution to (7, iii). If we set E = ∂1 u2 − ∂2 u1 − u3 − v3 , then, of course E(x1 , x2, 0) = 0, but if we compute ∂3 E = −{∂1 (u1 + v1 ) + ∂2 (u2 + v2 ) + ∂3 (u3 + v3 )}, we see that E vanishes identically if and only if the functions ua satisfy the additional equation (7, iv)

0 = ∂1 (u1 + v1 ) + ∂2 (u2 + v2 ) + ∂3 (u3 + v3 ).

This suggests modifying our Cauchy sequence by adjoining (7, iv), thus getting a new system with (s0 , s1 , s2 , s3 ) = (0, 1, 2, 0) and then proceeding as follows: (1*) Choose three functions ϕ1 (x1 ), ϕ2 (x1 , x2 ), ϕ3 (x1 , x2 ) (2*) Solve the equation in R2 , ∂2 w = −∂1 ϕ2 − ϕ3 − v¯3 with the initial condition w(x1 , 0) = ϕ1 (x1 ) where v¯3 (x1 , x2 ) = v3 (x1 , x2 , 0). (3*) Solve the triple of equations with initial conditions u1 (x1 , x2 , 0) = w(x1 , x2) ∂3 u1 = ∂1 u3 + u2 + v2 , u2 (x1 , x2 , 0) = ϕ2 (x1 , x2) ∂3 u2 = ∂2 u3 − u1 − v1 , ∂3 u3 = −∂1 (u1 + v1 ) − ∂2 (u2 + v2 ) − ∂3 v3 , u3 (x1 , x2 , 0) = ϕ3 (x1 , x2). It then follows easily that the resulting ua also satisfy (7, iii) (assuming that we are in the analytic category, so that we have existence and uniqueness in the Cauchy problem). In the example just given, the “compatibility condition” took the form of an extra equation which must be adjoined to the given equations so that the Cauchy sequence approach would work. One can imagine more complicated phenomena. Indeed, in the latter part of the nineteenth century, many examples of systems of P.D.E. were known to be tractable when treated as a sequence of initial value problems, provided that one was able to find a sufficient number of “compatibility conditions.” Around the turn of the century, Riquier [1910] and Cartan [1899] began to make a systematic study of this compatibility problem. Riquier’s approach was to work directly with the partial differential equations in question while Cartan, motivated by his research in differential geometry and Lie transformation groups (nowadays called Lie pseudo-groups), sought a coordinate-free approach. It was Cartan who realized that partial differentiation (which depends on a choice of coordinates) could be replaced by the exterior derivative operator (which does not). His method was to regard a collection of s functions ua of n variables xi as defining, via its graph, an n-dimensional submanifold of Rn+s . The condition that the collection ua satisfy a system of first order P.D.E. which was linear in the minors of the Jacobian matrix ∂u/∂x was then regarded as equivalent to the condition that the graph be an integral of a system S of differential forms on Rn+s . Cartan then let I be the differential ideal generated by the system S. The problem of constructing n-dimensional integral manifolds of I by a sequence of Cauchy problems then was reformulated as the problem of “extending” a p-dimensional integral manifold of I to one of dimension (p + 1).

56


Cartan’s next major insight into this problem was to realize that the condition that a submanifold N ⊂ M be an integral manifold of a differential ideal I on M is a condition only on the tangent planes of N . This led him to define the fundamental concept of integral elements of a differential system. Namely, the integral elements of dimension p of I are the p-planes E ⊂ Tz M on which all of the forms in I vanish. These form a closed subspace Vp (I) of Gp(T M ). Cartan’s approach was to study the structure of these subspaces and their interrelationships as p varies. Two important concepts arise which depend only on the structure of I as an algebraic ideal. These are the notions of ordinarity and regularity which are treated in detail in Section 1 of this Chapter. Roughly speaking, these concepts describe the smoothness of the spaces Vp (I) and the incidence spaces Vp,p+1 (I) ⊂ Vp (I) × Vp+1 (I). If one thinks intuitively of integral elements as “infinitesimal integral manifolds,” then these notions describe the well-posedness of the “infinitesimal Cauchy problem.” The main highlight of this section is Theorem 1.11, a version of Cartan’s test for an integral element to be ordinary. The version given here is an improvement over Cartan’s original version and was suggested to us by the recent work of W. K. Allard [1989]. In Section 2, after stating the classical Cauchy–Kowalevski theorem on the initial value problem for first order P.D.E., we state and prove the fundamental Cartan– K¨ ahler theorem. Roughly speaking, this theorem states that in the real-analytic category, the well-posedness of the initial value problem for an exterior differential ideal I is determined completely by the infinitesimal (algebraic) properties of the space of integral elements. Here, the condition that the ideal be differentially closed takes the place of the compatibility conditions which one must deal with in the P.D.E. formulation. We also discuss the classical terminology concerning the “generality” of the space of integral manifolds of a differential system, and introduce the important sequence of Cartan characters, which generalize the s0 , s1 , . . . , etc. described above. In Section 3, we consider a set of examples which demonstrate the use of the Cartan–K¨ ahler theorem in practice. Some of the examples are merely instructive while others are of interest in their own right. One example in particular, the isometric embedding example (Example 3.8), reproduces (with some improvements) Cartan’s original proof of the Cartan–Janet isometric embedding theorem. The following terminology will be used in the remainder of this chapter. If X is a smooth manifold and F ⊂ C ∞ (X) is any set of smooth functions on X, let Z(F ) ⊂ X denote the set of common zeros of the functions in F . We say that x ∈ Z(F ) is an ordinary zero of F if there exists a neighborhood V of x and a set of functions f 1 , f 2 , . . . , f q in F whose differentials are independent on V so that Z(F ) ∩ V = {y ∈ V | f 1 (y) = f 2 (y) = · · · = f q (y) = 0}. By the implicit function theorem, Z(F ) ∩ V is then a smooth submanifold of codimension q in V . Note that the set of ordinary zeroes of F is an open subset of Z(F ) (in the relative topology). If we let Z 0 (F ) denote the set of ordinary zeroes of F , then Z 0 (F ) is a disjoint union of connected, embedded submanifolds of X. Of course, the components of Z 0 (F ) do not all have to have the same dimension. By definition, the codimension of Z 0 (F ) at x ∈ Z 0 (F ) is the codimension in X of the component of Z 0 (F ) which contains x. A related piece of terminology is the following. If A ⊂ X is any subset and x ∈ A, then we say that A has codimension at most (resp., at least) q at x if there

§1. Integral Elements

57

exists an open neighborhood V of x ∈ X so that A ∩ V contains (resp., is contained in) a smooth (embedded) submanifold of V of codimension q which passes through x. Clearly A has codimension at least q at x and has codimension at most q at x if and only if A has the structure of a smooth submanifold of codimension q on a neighborhood of x.

§1. Integral Elements. Throughout this section, M will be a smooth manifold of dimension m and I ⊂ Ω∗ (M ) will be a differential ideal on M . Recall from Chapter I that an integral manifold of I is a submanifold ι : V → M with the property that ι∗ (α) = 0 for all α ∈ I. If v ∈ V and E = Tv V ⊂ Tv M is the tangent space to V at v, then ι∗ (α)|v = αE where, as usual, αE denotes the restriction of α|v to E ⊂ Tv M . It follows that the vanishing of ι∗(α)|v for all α ∈ I depends only on the tangent space of V at v. This leads to the following fundamental definition. Definition 1.1. Let M and I be as above. A linear subspace E ⊂ Tx M is said to be an integral element of I if ϕE = 0 for all ϕ ∈ I. The set of all integral elements of I of dimension p is denoted Vp (I). A submanifold of M is an integral manifold of I if and only if each of its tangent spaces is an integral element of I. Intuitively, one thinks of the integral elements of I as “infinitesimal integral manifolds” of I. It is not true, in general, that every integral element of I is tangent to an integral manifold of I. A simple counterexample is obtained by letting M = R1 and letting I be generated by the 1-form α = xdx. The space E = T0 R1 is an integral element of I, but E is clearly not tangent to any 1-dimensional integral manifold of I. A more subtle example (which will be used to illustrate several concepts in this section) is the following one. Example 1.2. Let M = R5 and let I be generated by the two 1-forms ϑ1 = dx1 + (x3 − x4 x5 )dx4 and ϑ2 = dx2 + (x3 + x4 x5 )dx5 . Then I is generated algebraically by the forms ϑ1 , ϑ2 , dϑ1 = ϑ3 ∧ dx4 , and dϑ2 = ϑ3 ∧ dx5 where we have written ϑ3 = dx3 + x5 dx4 − x4 dx5 . For each p ∈ M , let Hp = {v ∈ Tp R5 | ϑ1 (v) = ϑ2 (v) = 0} ⊂ Tp R5 . Then H ⊂ T R5 is a rank 3 distribution. A 1-dimensional subspace E ⊂ Tp R5 is an integral element of I if and only if E ⊂ Hp . Thus, V1 (I) ∼ = PH and it is a smooth manifold of dimension 7. Now let Kp = {v ∈ Tp R5 | ϑ1 (v) = ϑ2 v) = ϑ3 (v) = 0}. Then K ⊂ H is a rank 2 distribution on R5 . It is easy to see that, for each p ∈ R5 , Kp is the unique 2-dimensional integral element of I based at p. Thus, V2 (I) ∼ = R5 . Moreover, I has no integral elements of dimension greater than 2. It is not difficult to describe the 1-dimensional integral manifolds of I. Let f(t) = (f 3 (t), f 4 (t), f 5 (t)) be an arbitrary smooth immersed curve in R3 . There

58


exist functions f 1 (t), f 2 (t) (unique up to a choice of 2 constants) which satisfy the differential equations df 1 /dt = −(f 3 − f 4 f 5 )df 4 /dt df 2 /dt = −(f 3 + f 4 f 5 )df 5 /dt. Then F (t) = (f 1 (t), f 2 (t), f 3 (t), f 4 (t), f 5 (t)) is an integral manifold of I. Conversely, every 1-dimensional integral manifold of I is obtained this way. It is now easy to see that there exists an integral manifold of dimension 1 tangent to each element of V1 (I). On the other hand, by our calculation of V2 (I) above, any 2-dimensional integral manifold of I is an integral manifold of the differential system I+ generated by ϑ1 , ϑ2 , and ϑ3 . Using the fact that dϑ3 = −2 dx4 ∧ dx5 , we see that I+ is generated algebraically by ϑ1 , ϑ2 , ϑ3 , and dx4 ∧ dx5 . Hence I+ has no 2-dimensional integral elements, and a fortiori, no 2-dimensional integral manifolds. Thus, I has no 2dimensional integral manifolds either. As Example 1.2 shows, the relationship between the integral elements of a differential system and its integral manifolds can be subtle. In general, even the problem of describing the spaces Vn (I) can be complicated. The rest of this section will be devoted to developing basic properties of integral elements of I and of the subsets Vn (I). Proposition 1.3. If E is an n-dimensional integral element of I, then every subspace of E is also an integral element of I. Proof. Suppose that W ⊂ E is a subspace of E. If W were not an integral element of I, then there would be a form ϕ in I satisfying ϕW = 0. But then we would clearly have ϕE = 0, contradicting the assumption that E is an integral element of I. Proposition 1.4. Vn (I) = {E ∈ Gn (T M ) | ϑE = 0 for all ϑ in I of degree n}. Proof. The containment “⊂” is clear. Thus, we must prove that if ϑE = 0 for all ϑ in I of degree n then ϕE = 0 for all ϕ in I. Suppose that ϕE = 0 for some ϕ in I of degree p < n. Then there exists η0 in Λn−p(E ∗ ) so that ϕE ∧ η0 is a non-zero form in Λn (E ∗ ). Let η be a smooth (n − p)-form on M so that ηE = η0 . Then ϕ ∧ η is a form in I of degree n, but (ϕ ∧ η)E = ϕE ∧ η0 = 0. It follows from Proposition 1.4 that, for each x ∈ M , the set Vn (I) ∩ Gn (Tx M ) is an algebraic subvariety of Gn (Tx M ). The structure of this algebraic variety can be complicated. Fortunately, it is seldom necessary to confront this problem directly. In practice, the spaces Vn (I) are most often studied by an inductive procedure which uses information about Vp (I) to get information about Vp+1 (I). The ultimate reason for this approach will be clear in the next section when we prove the Cartan–K¨ ahler theorem, which builds integral manifolds of I by solving a sequence of initial value P.D.E. problems. A more immediate reason will be furnished by Proposition 1.6 below.


59

Definition 1.5. Let e1 , e2 , . . . , ep be a basis of E ⊂ Tx M . We define the polar space of E to be the vector space H(E) = {v ∈ Tx M | ϕ(v, e1 , e2 , . . . , ep ) = 0 for all ϕ in I of degree p + 1}. Note that E ⊂ H(E). The annihilator of H(E) is denoted E(E) ⊂ Tx∗ M and is referred to as the space of polar equations of E. The importance of H(E) is explained by the following proposition. Proposition 1.6. Let E be an integral element of I of dimension p. Then a (p + 1)-plane E + containing E is an integral element of I if and only if it satisfies E + ⊂ H(E). Proof. Suppose that E + = E + Rv and that e1 , e2 , . . . , ep is a basis of E. By Proposition 1.4, E + is an integral element of I if and only if ϕE + = 0 for all (p + 1)-forms ϕ in I. By definition, this latter condition holds if and only if v lies in H(E). Even though the space Vp+1 (I) ∩ Gp+1 (Tx M ) may be a complicated algebraic variety, for a fixed E ∈ Vp (I), the space of those E + ∈ Vp+1 (I) which contain E is a (real) projective space which is canonically isomorphic to P(H(E)/E). This motivates us to define a function r : Vp (I) → Z by the formula r(E) = dim H(E) − (p + 1). Note that r(E) ≥ −1 with equality if and only if E lies in no (p + 1)-dimensional integral element of I, i.e., E is maximal. When r(E) ≥ 0, the set of (p + 1)dimensional integral elements of I which contain E is then a real projective space of dimension r(E). This linearization of an exterior algebra problem is related to the linearization process in multi-linear algebra known as “polarization,” the most common example being the polarization of a quadratic form on a vector space to produce a bilinear form. We shall not try to make this relationship more precise. We merely offer this comment as a motivation for the name “polar space” for H(E), which was ´ Cartan. Other authors have referred to H(E) as the “space of integral coined by E. enlargements of E” or used similar terminology. If Ω is any n-form on M , let Gn (T M, Ω) denote the open set consisting of those E’s for which ΩE = 0. If ϕ is any other n-form on M , we can define a function ϕΩ on Gn (T M, Ω) by the formula ϕE = ϕΩ (E)ΩE for all E ∈ Gn (T M, Ω). (Since Λn (E ∗ ) is 1-dimensional with basis ΩE , this definition makes sense.) By Proposition 1.4, the set Vn (I, Ω) = Vn (I) ∩ Gn (T M, Ω) is the space of common zeroes of the set of functions FΩ (I) = {ϕΩ | ϕ lies in I and has degree n}. Definition 1.7. An integral element E ∈ Vn (I) will be said to be K¨ ahler-ordinary if there exists an n-form Ω on M with ΩE = 0 with the property that E is an ordinary zero of the set of functions FΩ (I). We shall use the notation Vno (I) ⊂ Vn (I) to denote the subspace of K¨ ahler-ordinary points of Vn (I). If E is a K¨ ahler-ordinary integral element and the function r is locally constant on a neighborhood of E in

60


Vno (I), then we say that E is K¨ ahler-regular. We shall use the notation Vnr (I) ⊂ o Vn (I) to denote the subspace of K¨ ahler-regular points of Vno (I). The role of Ω in the above definition is not critical. If Ω and Ψ are two n-forms with ΩE = 0 and ΨE = 0, then E ∈ Gn (T M, Ω) ∩ Gn (T M, Ψ) and the identity ϕΩ = ϕΨ · ΨΩ holds on Gn (T M, Ω) ∩ Gn (T M, Ψ). Since ΨΩ never vanishes on Gn (T M, Ω) ∩ Gn (T M, Ψ), it follows that E is an ordinary zero of FΩ (I) if and only if it is an ordinary zero of FΨ (I). Note that Vno (I) is an embedded submanifold of Gn (T M ) and is an open subset of Vn (I) in the relative topology. Since r is an upper semicontinuous function on Vno (I), Vnr (I) is a open, dense subset of Vno (I). Example 1.2 (continued). We will show that all of the 2-dimensional integral elements of I are K¨ ahler-regular. Let Ω = dx4 ∧ dx5 . Then every element E ∈ 5 G2 (T R , Ω) has a unique basis of the form X4 (E) = ∂/∂x4 + p14 (E)∂/∂x1 + p24 (E)∂/∂x2 + p34 (E)∂/∂x3 X5 (E) = ∂/∂x5 + p15 (E)∂/∂x1 + p25 (E)∂/∂x2 + p35 (E)∂/∂x3 . The functions x1 , . . . , x5, p14 , . . . , p35 form a coordinate system on G2 (T R5 , Ω). Computation gives (ϑ1 ∧ dx4 )Ω = −p15 (ϑ1 ∧ dx5 )Ω = p14 + (x3 − x4 x5 ) (ϑ2 ∧ dx4 )Ω = −p25 − (x3 + x4 x5 ) (ϑ2 ∧ dx5 )Ω = p24 (ϑ3 ∧ dx4 )Ω = −p35 + x4 (ϑ3 ∧ dx5 )Ω = +p34 + x5 . These 6 functions are clearly independent on G2 (T R5 , Ω) and their common zeroes are exactly V2 (I). Thus, every point of V2 (I) is K¨ ahler-ordinary. Since none of these elements has any extension to a 3-dimensional integral element, it follows that r(E) = −1 for all E ∈ V2 (I). Thus, every element of V2 (I) is also Kähler-regular. Similarly, it is easy to see that every E ∈ V1 (I) is K¨ ahler-ordinary. However, not every element of V1 (I) is K¨ ahler-regular. To see this, note that any E ∈ V1 (I) on which ϑ3 does not vanish cannot lie in any 2-dimensional integral element of I. Thus, r(E) = −1 for all E ∈ V1 (I, ϑ3). On the other hand, each E + ∈ V1 (I) on which ϑ3 does vanish lies in a unique E + ∈ V2 (I) and hence has r(E) = 0. Since V1 (I, ϑ3) is clearly dense in V1 (I), it follows that V1r (I) = V1 (I, ϑ3 ). Returning to the general theory, we shall need to understand the following incidence correspondences: Vp,p+1 (I) = {(E, E + ) ∈ Vp (I) × Vp+1 (I) | E ⊂ E + } r (I) = {(E, E + ) ∈ Vpr (I) × Vp+1 (I) | E ⊂ E + }. Vp,p+1

We let πp : Vp,p+1 (I) → Vp (I) denote the projection onto the first factor and we let πp+1 : Vp,p+1 (I) → Vp+1 (I) denote the projection onto the second factor.


61

The fibers of these maps are easy to describe. If E ∈ Vp (I) has r(E) ≥ 0, then (πp )−1 (E) ∼ = P(H(E)/E) ∼ = RPr(E) . On the other hand, if E + ∈ Vp+1 (I), then −1 + ∼ + ∗ (πp+1 ) (E ) = P(E ) , the space of hyperplanes in E + . It is helpful to keep in mind the following diagram. Vp,p+1 (I) πp+1 Vp (I) Vp+1 (I) πp

This “double fibration” fails, in general, to be surjective or submersive on either r base. The next proposition shows that the picture is better for Vp,p+1 (I). Its proof, although technical, is straightforward. Some care is needed to prove that the regularity assumption suffices to guarantee that certain maps have maximal rank. r (I) is not empty, then it is a smooth manifold. MoreProposition 1.8. If Vp,p+1 r o (I) and both of the maps over, the image πp+1 (Vp,p+1 (I)) is an open subset of Vp+1 r r r r (I)) are submersions. πp : Vp,p+1 (I) → Vp (I) and πp+1 : Vp,p+1 (I) → πp+1 (Vp,p+1

Proof. Let E ∈ Vpr (I) have base point z ∈ M and let t = dim M − dim H(E). By hypothesis, there exist t (p + 1)-forms κ1 , κ2 , . . . , κt in I so that, for any basis e1 , e2 , . . . , ep of E, we have H(E) = {v ∈ Tz M | κτ (v, e1 , e2 , . . . , ep ) = 0 for 1 ≤ τ ≤ t}. Since r is locally constant on a neighborhood of E in Vpr (I), it follows easily that we must have ˜ = {v ∈ Tz˜M | κτ (v, e˜1 , e˜2 , . . . , e˜p ) = 0 for 1 ≤ τ ≤ t} H(E) for all E˜ in Vpr (I) (based at z˜, with basis e˜1 , e˜2 , . . . , e˜p ) sufficiently near E. From this, it follows that if we set H = {(E, v) ∈ Vp (I) × T M | v ∈ H(E)}, then H is a family of vector spaces over Vp (I) which restricts to each component of Vpr (I) to be a smooth vector bundle of constant rank. We also conclude that, for each component Z of Vpr (I) on which r is non-negative, the component (πp )−1 (Z) of r (I) is a smooth bundle over Z. Vp,p+1 r (I) is open in Vp+1 (I). We now show that the image of πp+1 restricted to Vp,p+1 + r Let (E, E ) belong to Vp,p+1 (I). There exists an open neighborhood of E, U ⊂ Gp (T M ), so that U ∩ Vp (I) ⊂ Vpr (I) and an open neighborhood of E + , U + ⊂ ˜ in U . Thus, if E˜ + ∈ Gp+1 (T M ), so that every E˜ + in U + contains a p-plane E + r r ˜ (I)) contains U ∩Vp+1 (I), then E ∈ U ∩Vp (I) ⊂ Vp (I). It follows that πp+1 (Vp,p+1 + U ∩ Vp+1 (I). r o It remains to show that πp+1 (Vp,p+1 (I)) lies in Vp+1 (I) and that πp+1 restricted r to Vp,p+1 (I) is a submersion onto its image. To do this, we choose coordinates. Let r (I), let r = r(E) ≥ 0, and let t = dim M − dim H(E) = (E, E + ) belong to Vp,p+1 dim M − (r + p + 1). The cases where either r or t are zero can be handled by obvious simplifications of the following argument, so we assume that r and t are positive. Choose coordinates x1 , . . . , xp , y, v1 , . . . , vr , u1 , . . . , ut centered on the base point z of E with the properties (i) E is spanned by the vectors ∂/∂xi at z for 1 ≤ i ≤ p.

62


(ii) E + is spanned by the vectors in E and the vector ∂/∂y. (iii) H(E) is spanned by the vectors in E + and the vectors ∂/∂vρ at z for 1 ≤ ρ ≤ r. Let Ω = dx1 ∧ · · · ∧ dxp. By hypothesis, there exist a set of p-forms {ϕ1 , . . . , ϕq } in I and an E-neighborhood U ⊂ Gp (T M, Ω) so that the functions f c = ϕcΩ for 1 ≤ c ≤ q have independent differentials on U and ˜ ∈ U | f c (E) ˜ = 0 for all c}. Vp (I) ∩ U = {E We may also suppose, by shrinking U if necessary, that there are t (p + 1)-forms κ1 , κ2 , . . . , κt in I so that ˜ = {v ∈ Tz˜M | κτ (v, e˜1 , e˜2 , . . . , e˜p ) = 0 for 1 ≤ τ ≤ t} H(E) for all E˜ in Vp (I) ∩ U (based at z˜, with basis e˜1 , e˜2 , . . . , e˜p ). We now want to show that if we set Ω+ = Ω ∧ dy and define gc = (ϕc ∧ dy)Ω+ and hτ = (κτ )Ω+ then there is an open neighborhood U + ⊂ Gp+1 (T M, Ω+ ) of E + so that the set of functions {gc , hτ } have independent differentials on U + and that ˜ + ∈ U + | gc (E˜ + ) = hτ (E˜ + ) = 0 for all c and τ }. Vp+1 (I) ∩ U + = {E o (I). In particular, we will conclude that E + ∈ Vp+1 + + ˜ Note that every E ∈ Gp+1 (T M, Ω ) contains a unique p-plane, which we will ˜ ⊂ E˜ + , on which the differential dy vanishes. Let U + ⊂ Gp+1 (T M, Ω+ ) be denote E + an E -neighborhood in Gp+1 (T M ) so that E˜ ∈ U whenever E˜ + ∈ U + . There exist unique functions Aρi , Biτ , aρ , and bτ on Gp+1 (T M, Ω+ ) so that the p + 1 vectors

˜ + ) = ∂/∂xi + Aρ (E˜ + )∂/∂vρ + B τ (E˜ + )∂/∂uτ Xi (E i i ˜ + ) = ∂/∂y + aρ (E˜ + )∂/∂vρ + bτ (E ˜ + )∂/∂uτ Y (E ˜ Note that the functions are a basis of E˜ + . The vectors Xi (E˜ + ) form a basis of E. x, y, v, u, A, B, a, and b form a coordinate system on Gp+1 (T M, Ω+ ). ˜ for all c and all Since dy(Xi (E˜ + )) = 0, we have the formula gc (E˜ + ) = f c (E) + + + + c ˜+ ˜ ˜ E ∈ Gp+1 (T M, Ω ). It follows that, if E ∈ U and g (E ) = 0 for all c, then ˜ + ), Y (E˜ + )), it follows that E˜ ∈ Vp (I) ∩ U . Since hτ (E˜ + ) = κτ (X1 (E˜ + ), . . . , Xp (E ˜ whenever E ˜ ∈ Vp (I) ∩ U . the equations hτ (E˜ + ) = 0 imply that Y (E˜ + ) lies in H(E) It follows that ˜ + ∈ U + | gc (E˜ + ) = hτ (E˜ + ) = 0 for all c and τ }. Vp+1 (I) ∩ U + ⊃ {E Since the reverse inclusion is clear, we have proved equality. It remains to show that the functions {gc , hτ } have linearly independent differen˜ + )), tials at E + . To see this, first note that since hτ (E˜ + ) = κτ (X1 (E˜ + ), . . . , Xp (E˜ + ), Y (E τ when we expand the functions h in terms of the coordinates (x, y, v, u, A, B, a, b) they are linear in the functions {aρ , bτ }. Thus hτ = N τ + Mντ bν for some coefficients N and M that depend only on (x, y, v, u, A, B, a). By hypothesis, we have Nρτ (E + ) = 0 and the t×t matrix (Mντ (E + )) is invertible. It follows that, by shrinking U + if necessary, we may suppose that (Mντ (E˜ + )) is invertible for all E˜ + ∈ U + .


63

Hence we may write hτ = Mντ (bν − T ν ) where the functions T ν depend only on the variables x, y, v, u, A, B and a. It follows that the functions {gc , hτ } have independent differentials at E + if and only if the functions {gc , bτ − T τ } have independent differentials at E + . Since the functions gc can be expressed in terms of the coordinates x, y, v, u, A, and B alone, it follows that the functions {gc , bτ − T τ } have independent differentials at E + if and only if the functions {gc } have independent differentials at E + . Let K ⊂ U be the set of p-planes on which the differential dy vanishes. Then K is clearly a smooth submanifold of U which contains E. Since E˜ ˜ + ) = f c (E), ˜ ˜ + lies in U + , and since we have the identity gc (E lies in K whenever E c + it follows that functions {g } have independent differentials at E if and only if the functions {f c } have independent differentials at E after they have been restricted to K. Now every E˜ ∈ U has a unique basis of the form ρ τ ˜ = ∂/∂xi + wi (E)∂/∂y ˜ ˜ ˜ Xi (E) + Aρi (E)∂/∂v + Biτ (E)∂/∂u ,

and the functions x, y, v, u, A, B, and w form a coordinate system on U centered on ˜ ∈ U | wi (E) ˜ = 0 for all i}. It follows that the functions E. Also, we have K = {E c {f } have independent differentials at E after they have been restricted to K if and only if the functions {wi } have independent differentials on the set Vp (I) ∩ U = ˜ = 0 for all c}. However, since E + ∈ Vp+1 (I), it follows that, for ˜ ∈ U | f c (E) {E any vector λ = (λ1 , . . . , λn ) where the λi are sufficiently small, Vp (I) ∩ U contains the p-plane Eλ ⊂ E + which is spanned by the vectors Xi (λ) = ∂/∂xi + λi ∂/∂y for i between 1 and p. Since the functions {wi } are independent when restricted to the p-manifold E = {Eλ | Eλ ∈ U } ⊂ Vp (I) ∩ U , we are done. Since we have shown that E + is a Kähler-ordinary integral element of I, it r o follows that πp+1 (Vp,p+1 (I)) is an open subset of Vp+1 (I). The fact that πp+1 is a r submersion when restricted to Vp,p+1 (I) is now elementary. r The proof of Proposition 1.8 has an important corollary: If (E, E + ) ∈ Vp,p+1 (I), then the following formula holds

(8)

(codim Vp+1 (I) in Gp+1 (T M ) at E + ) = (codim Vp (I) in Gp (T M ) at E) + (codim H(E) in Tz M ).

A nested sequence of subspaces (0)z ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ Tz M where each Ek is of dimension k and En is an integral element of I is called an integral flag of I of length n based at z. If z is an ordinary point of the set of functions F = I ∩ Ω0 (M ) (i.e., the set of 0-forms in I), and the function r is locally constant on a neighborhood of Ek in Vk (I) for all k ≤ n − 1, then Proposition 1.8 applies inductively to show that each Ek is Kähler-regular for k ≤ n − 1 and that En is K¨ ahler-ordinary. Definition 1.9. Let I be a differential system on a manifold M . An integral element E ∈ Vn (I) is said to be ordinary if its base point z ∈ M is an ordinary zero of I ∩ Ω0 (M ) and moreover there exists an integral flag (0)z ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ Tz M with E = En where Ek is Kähler-regular for k ≤ n − 1. If E is both ordinary and K¨ ahler-regular, then we say that E is regular. In an integral flag (0)z ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ Tz M where each Ek is Kählerregular for k ≤ n − 1, note that each Ek is actually regular for k ≤ n − 1. Such

64


a flag is called an ordinary flag. If, in addition, En is also regular, the flag is said to be regular. Note that, for integral elements, we have the implications: regular ⇒ K¨ ahler-regular, ordinary ⇒ K¨ ahler-ordinary, and regular ⇒ ordinary. However, these implications are not generally reversible. Our goal in the remainder of this section is to describe one of the fundamental tests for an integral element to be ordinary. First, we shall introduce a set of constructions which are frequently useful in the study of integral flags of I. It is convenient to assume that the differential ideal I contains no non-zero forms of degree 0. Since this is usually the case in practice, this restriction is not unreasonable. Proposition 1.10. Let I ⊂ Ω+ (M ) be a differential ideal which contains no nonzero forms of degree 0. Let (0)z ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ Tz M be an integral flag of I. Let e1 , e2 , . . . , en be a basis of En so that e1 , e2 , . . . , ek is a basis of Ek for all 1 ≤ k ≤ n. For each k ≤ n, let ck be the codimension of H(Ek ) in Tz M . The numbers ck satisfy ck−1 ≤ ck . For each integer a between 1 and cn−1 , define the level of a, denoted λ(a), to be the smallest integer so that a ≤ cλ(a) . If cn−1 > 0, then there exists a sequence ϕ1 , . . . , ϕcn−1 of forms in I so that ϕa has degree λ(a) + 1 and so that for all 0 ≤ k ≤ n − 1, (9)

H(Ek ) = {v ∈ Tz M | ϕa (v, e1 , . . . , eλ(a)) = 0 for all a ≤ ck }.

Proof. Since it is clear that H(Ek+1 ) ⊂ H(Ek ), it follows that ck+1 ≥ ck . To construct the sequence ϕ1 , . . . , ϕcn−1 , we proceed by induction on the level k. By the very definition of H(E0 ), there exist 1-forms ϕ1 , . . . , ϕc0 in I so that (9) holds for k = 0. Suppose now that we have constructed a sequence ϕ1 , . . . , ϕcp−1 so that (9) holds for all k < p. Let ω1 , . . . , ωn be a sequence of 1-forms on M so that their restriction to E is the dual coframe to e1 , e2 , . . . , en . Define ϕã ∈ I by ϕã = ϕa ∧ ωλ(a)+1 ∧ ωλ(a)+2 ∧ · · · ∧ ωp . Then ϕ ã is a form in I of degree p + 1 and, a since ϕ vanishes on E, the identity (10)

ϕã (v, e1 , . . . , ep ) = ϕa (v, e1 , . . . , eλ(a))

holds for all v ∈ Tz M . If cp = cp−1 , then H(Ep ) = H(Ep−1 ), so (10) shows that (9) already holds for k = p. If cp > cp−1 , then by the definition of H(Ep ), we can choose a set of (p + 1)-forms in I, {ϕa | cp−1 < a ≤ cp }, so that H(Ep) is the set of vectors v satisfying ϕã (v, e1 , . . . , ep ) = 0 for all a ≤ cp−1 as well as ϕa (v, e1 , . . . , ep ) = 0 for all cp−1 < a ≤ cp . This completes the induction step. A sequence ϕ1 , . . . , ϕcn−1 of forms in I with the properties given in Propostion 1.10 will be call a polar sequence associated to the integral flag (0)z ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ Tz M . Note that the polar sequence does not necessarily carry complete information about H(En ). Theorem 1.11 (Cartan’s test). Let I ⊂ Ω+ (M ) be an ideal which contains no non-zero forms of degree 0. Let (0)z ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ Tz M be an integral flag of I and, for each k < n, let ck be the codimension of H(Ek ) in Tz M . Then Vn (I) ⊂ Gn (T M ) is of codimension at least c0 + c1 + · · · + cn−1 at En . Moreover, each Ek is regular for all k < n (and hence En is ordinary) if and only if En has a neighborhood U in Gn (T M ) so that Vn (I) ∩ U is a smooth manifold of codimension c0 + c1 + · · · + cn−1 in U .


65

Proof. Set s = dim M − n. There exists a z-centered local coordinate system x1 , . . . , xn , u1 , . . . , us with the property that Ek is spanned by the vectors {∂/∂xi }i≤k and so that, for all k < n, H(Ek ) = {v ∈ Tz M | dua (v) = 0 for all a ≤ ck }. Let ϕ1 , . . . , ϕcn−1 be a polar sequence for the given flag so that dua (v) = ϕa (v, ∂/∂x1 , ∂/∂x2 , . . . , ∂/∂xλ(a) ) for all v ∈ Tz M . It follows that ϕa = dua ∧ dx1 ∧ dx2 ∧ · · · ∧ dxλ(a) + ψa , where ψa is a form of degree λ(a) + 1 which is a sum of terms of the following three kinds: (i) dub ∧ dxJ where J is a multi-index of degree λ(a) which contains at least one index j which is larger than λ(a), (ii) forms which vanish at z, (iii) forms which are of degree at least 2 in the differentials {dub }. We are now going to show that the forms {ϕa | a ≤ cn−1 } suffice to generate a set of at least c0 + c1 + · · ·+ cn−1 functions on a neighborhood of En whose differentials are linearly independent at En and whose set of common zeroes contains Vn (I) in a neighborhood of En . Let Ω = dx1 ∧ dx2 ∧ · · · ∧ dxn . Let Gn (T U, Ω) ⊂ Gn (T M ) be the set of n-planes E which are based in the domain U of the (x, u)-coordinates and for which ΩE = 0. Then there exist functions {pai | 1 ≤ j ≤ n and 1 ≤ a ≤ s} on Gn (T U, Ω) so that, for each E ∈ Gn (T U, Ω), the vectors Xi (E) = ∂/∂xi + pai(E)∂/∂ua form a basis of E. The functions (x, u, p) form an En -centered coordinate system on Gn (T U, Ω). For convenience, we define λ(a) = n for all cn−1 < a ≤ s. Let us say that a pair of integers (j, a) where 1 ≤ j ≤ n and 1 ≤ a ≤ s is principal if it satisfies j ≤ λ(a) otherwise, we say that the pair is non-principal. (For example, there are no principal pairs if c0 = s.) Since, for j ≥ 1, there are cj − cj−1 values of a in the range 1 ≤ a ≤ s which satisfy λ(a) = j, it easily follows that the number of principal pairs is ns − (c0 + c1 + · · · + cn−1 ). Hence, the number of non-principal pairs is c0 + c1 + · · · + cn−1 . Let (j, a) be a non-principal pair. Define the function Fja on Gn (T U, Ω) by a Fj (E) = ϕa (Xj (E), X1 (E), X2 (E), . . . , Xλ(a) (E)). Then we have an expansion Fja = paj + Pja + Qaj where Pja is a linear combination (with constant coefficients) of the variables xi , ua , and {pai | (i, a) is principal} and Qaj vanishes to second order at En . This expansion follows directly from an examination of the terms in ψa as described above. It follows that the functions {Fja | (j, a) is non-principal} have linearly independent differentials at En . Let U ⊂ Gn (T U, Ω) be a neighborhood of En on which these functions have everywhere linearly independent differentials. Then we clearly have Vn (I) ∩ U ⊂ {E ∈ U | Fja (E) = 0 for all non-principal (j, a)}.

66


It follows that Vn (I) has codimension at least c0 + c1 + · · · + cn−1 at En , as desired. This proves the first part of the theorem. In order to prove the second statement of the theorem, we begin by supposing that each Ek is K¨ ahler-regular for all 1 ≤ k < n. Then, by definition, En is ordinary. Then (8) shows that we have the following recursion formula for all k between 1 and n: (codim Vk (I) in Gk (T M ) at Ek ) = ck−1 + (codim Vk−1 (I) in Gk−1 (T M ) at Ek−1 ). Since, by hypothesis, I contains no 0-forms, it follows that V0 (I) = G0 (T M ) = M . Thus, by induction, the codimension of Vn (I) in Gn (T M ) at En has the desired value c0 + c1 + · · · + cn−1 . To prove the converse statement, let us now suppose that there is an En neighborhood U in Gn (T M ) so that Vn (I)∩ U is a smooth manifold of codimension c0 +c1 +· · ·+cn−1 in U . It follows that, by shrinking U if necessary, we may suppose that Vn (I) ∩ U = {E ∈ U | Fja (E) = 0 for all non-principal (j, a)} and that the functions F = {Fja | (j, a) is non-principal} have linearly independent differentials on all of U . If we set ϕaj = ϕa ∧ dxK(a,j) where K(a, j) is a multi-index of degree n − (λ(a) + 1) with the property that dxj ∧ dx1 ∧ dx2 ∧ · · · ∧ dxλ(a) ∧ dxK(a,j) = dx1 ∧ dx2 ∧ · · · ∧ dxn , then Fja (E) = ϕaj (X1 (E), X2 (E), . . . , Xn (E)) for all E ∈ U . It follows that En is K¨ ahler-ordinary. In fact, we can say more. Applying the implicit function theorem to the above expansion of Fja , it follows that, by shrinking U if necessary, the submanifold Vn (I) ∩ U in U can be described by equations of the form pai = Pia for all (i, a) non-principal, where the functions Pja are functions of the variables xi , ua , and {pai | (i, a) is principal}. (For the sake of uniformity, we define the function Pia to be pai when the pair (i, a) is principal.) Thus, the variables xi , ua , and {pai | (i, a) is principal} form an En -centered coordinate system on Vn (I) ∩ U . By the first part of the proof, we know that Vn−1 (I) has codimension at least c0 + c1 + · · · + cn−2 in Gn−1 (T M ) at En−1 . We will now show that, in fact, Vn−1 (I) contains a submanifold of codimension c0 +c1 +· · ·+cn−2 in Gn−1 (T M ) which passes through En−1. This will imply that En−1 is Kähler-ordinary and that Vn−1 (I) has codimension c0 + c1 + · · · + cn−2 in Gn−1(T M ) at En−1. To demonstrate this claim, let v = (v1 , v2 , . . . , vn−1 ) ∈ Rn−1 , and define a map Φ : Rn−1 ×Vn (I)∩U → Vn−1 (I) by letting Φ(v, E) = E v be the (n − 1)-plane in E which is spanned by the n − 1 vectors Xi (E v ) = Xi (E) + vi Xn (E)

for all 1 ≤ i ≤ n − 1

= ∂/∂xi + vi ∂/∂xn + (Pia (E) + vi Pna(E))∂/∂ua . We claim that Φ has rank ρ = n + s + (n − 1)(s + 1) − (c0 + c1 + · · · + cn−2 ) at (0, En ). To see this, note that since c0 +c1 +· · ·+cn−2 is already known to be a lower


67

bound on the codimension of Vn−1 (I) in Gn−1 (T M ) at En−1 , the image of Φ must lie in a submanifold of Gn−1 (T M ) whose codimension is at least c0 + c1 + · · ·+ cn−2 and hence the rank of Φ cannot be larger than ρ at any point of some neighborhood of (0, En ). On the other hand, the rank of Φ at (0, En) is equal to ρ, since it is clear that the ρ functions xi , ua, {vi | 1 ≤ i ≤ n − 1}, and {Pia (E) + vi Pna (E) | (i, a) principal and i ≤ n − 1} have linearly independent differentials on a neighborhood of (0, En ). Thus, the rank of Φ must be identically ρ near (0, En). Moreover, there is a neighborhood O of (0, En ) in Rn−1 × Vn (I) ∩ U and a neighborhood U − of En−1 in Gn−1 (T M ) so that Vn−1 (I) ∩ U − is a smooth submanifold of U − of codimension c0 + c1 + · · · + cn−2 and so that Φ : O → Vn−1 (I) ∩ U − is a surjective submersion. As noted above, this implies that En−1 is Kähler-ordinary. We may also conclude that En−1 is Kähler-regular by the following observation. ˜ for some v} is an open For all E˜ in U −, the set {E ∈ Vn (I) ∩ U | Φ(v, E) = E ˜ ˜ ˜ subset of the set P(H(E)/E) of n-dimensional integral elements which contain E. ˜ The dimension of this set is thus r(E). However, since Φ is a surjective submersion, this set clearly has the dimension of the fibers of Φ, which is the same as the dimension of the fiber Φ−1 (En−1). It follows that the function r is locally constant on a neighborhood of En−1 in Vn−1 (I). Thus, En−1 is Kähler-regular, as desired. By induction, it follows that each Ek is Kähler-regular for all 1 ≤ k ≤ n − 1. Since I contains no forms of degree 0, it immediately follows that each Ek is regular for each 1 ≤ k ≤ n − 1. Example 1.2 (continued). Using Theorem 1.11, we can give a quick proof that none of the elements in V2 (I) are ordinary. For any integral flag (0)z ⊂ E1 ⊂ E2 ⊂ Tz R5 , we know that c0 ≤ 2 since there are only 2 independent 1-forms in I. Also, since E2 ⊂ H(E1 ), it follows that c1 ≤ 3. Since there is a unique 2-dimensional integral element at each point of R5 , it follows that V2 (I) has codimension 6 in G2 (T R5 ). Since 6 > c0 +c1 , it follows, by Theorem 1.11, that none of the integral flags of length 2 can be ordinary. Hence there are no ordinary integral elements of dimension 2. Example 1.12. Let M = R6 with coordinates x1 , x2 , x3 , u1 , u2 , u3. Let I be the differential system generated by the 2-form ϑ = d(u1 dx1 + u2 dx2 + u3 dx3 ) − (u1 dx2 ∧ dx3 + u2 dx3 ∧ dx1 + u3 dx1 ∧ dx2 ). Of course, I is generated algebraically by the forms {ϑ, dϑ}. We have dϑ = −(du1 ∧ dx2 ∧ dx3 + du2 ∧ dx3 ∧ dx1 + du3 ∧ dx1 ∧ dx2 ). We can use Theorem 1.11 to show that all of the 3-dimensional integral elements of I on which Ω = dx1 ∧ dx2 ∧ dx3 does not vanish are ordinary. Let E ∈ V3 (I, Ω) be fixed with base point z ∈ R6 . let (e1 , e2 , e3 ) be the basis of E which is dual to the basis (dx1 , dx2 , dx3) of E ∗ . Let E1 be the line spanned by e1 , let E2 be the 2-plane spanned by the pair {e1 , e2 }, and let E3 be E. Then (0)z ⊂ E1 ⊂ E2 ⊂ E3 is an integral flag. Since I is generated by {ϑ, dϑ} where ϑ is a 2-form, it follows that c0 = 0. Moreover, since ϑ(v, e1 ) = π1 (v) where π1 ≡ du1 mod (dx1 , dx2, dx3 ), it follows that c1 = 1. Note that, since H(E2 ) ⊃ E3 , it follows that c2 ≤ 3. On the other hand, we have the formula ϑ(v, e1 ) = π1 (v) ϑ(v, e2 ) = π2 (v) dϑ(v, e1 , e2 ) = −π3 (v)

68


where in each case, πk ≡ duk mod (dx1 , dx2, dx3). Since the 1-forms πk are clearly independent and annihilate H(E2 ), it follows that c2 ≥ 3. Combined with the previous argument, we have c2 = 3. It follows by Theorem 1.11 that the codimension of V3 (I) in G3 (T R6 ) at E is at least c0 + c1 + c2 = 4. We are now going to show that V3 (I, Ω) is a smooth submanifold of G3 (T R6 ) of codimension 4, and thence, by Theorem 1.11, conclude that E is ordinary. To do this, we introduce functions pij on G3 (T R6 , Ω) with the property that, for each E ∈ G3 (T R6 , Ω) based at z ∈ R6 , the forms πi = dui − pij (E)dxj ∈ Tz∗ (R6 ) are a basis for the 1-forms which annihilate E. Then the functions (x, u, p) form a coordinate system on G3 (T R6 , Ω). It is easy to compute that ϑE = (p23 − p32 − u1 )dx2 ∧ dx3 + (p31 − p13 − u2 )dx3 ∧ dx1 + (p12 − p21 − u3 )dx1 ∧ dx2 dϑE = −(p11 + p22 + p33 )dx1 ∧ dx2 ∧ dx3 . It follows that the condition that E ∈ G3 (T R6 , Ω) be an integral element of I is equivalent to the vanishing of 4 functions on G3 (T R6 , Ω) whose differentials are independent. Thus V3 (I, Ω) is a smooth manifold of codimension 4 in G3 (T R6 , Ω), as we desired to show. The following results will be used in later sections: Proposition 1.13. Let I ⊂ Ω∗ (M ) be a differential ideal which contains no nonzero forms of degree 0. Let Z ⊂ Vn (I) be a connected component of the space of ordinary integral elements. Then there exists a unique sequence (c0 , c1 , c2 , . . . , cn−1) of integers so that ck is the codimension of H(Ek ) in Tz M for any ordinary integral flag (0) ⊂ E1 ⊂ · · · ⊂ En ⊂ Tz M with En ∈ Z. r (I) × Z denote the space of ordinary Proof. Let Z˜ ⊂ V0r (I) × V1r (I) × . . . × Vn−1 integral flags F = (E0 , E1, . . . , En ) of I with En ∈ Z. We endow Z˜ with the topology and smooth structure it inherits from this product. Note that even though Z is connected, Z˜ may not be connected. However, if we define ck (F ) = dim M − ˜ We dim H(Ek ), then the functions ck for k < n are clearly locally constant on Z. ˜ must show that these functions are actually constant on Z. ˜ Then there To do this, suppose that for some p < n, cp were not constant on Z. would exist non-empty open sets Z˜1 , Z˜2 so that cp ≡ q on Z˜1 and cp = q on Z˜2 . The images Z1 , Z2 of these two sets under the submersion Z˜ → Z would then be an open cover of Z. By the connectedness of Z, they would have to intersect non-trivially. In particular, there would exist an E ∈ Z and two p-planes E 1 , E 2 ∈ Vpr (I)∩ Gp(E) for which r(E 1 ) = q = r(E 2 ). We shall now show that this is impossible. Since E ⊂ Tz M is an integral element, it follows that Gp (E) ⊂ Vp (I) and hence that Vpr (I) ∩ Gp (E) is an open subset of Gp (E). Moreover, since the function r is locally constant on Vpr (I), it follows that Vpr (I) ∩ Gp (E) is a subset of the open set G∗p (E) ⊂ Gp(E) on which r is locally constant. Thus, it suffices to show that r is constant on G∗p (E). Let ϕ1 , . . . , ϕq be a set of (p+1)-forms in I with the property that a (p+1)-plane Ep+1 ⊂ Tz M is an integral element of I if and only if each of the forms ϕ1 , . . . , ϕq vanish on Ep+1 . (Since we are only considering planes based at z, such a finite


69

collection of forms exists.) Then for any Ep ∈ Gp (E) with basis e1 , e2 , . . . , ep , H(Ep ) = {v ∈ Tz M | ϕa (v, e1 , e2 , . . . , ep ) = 0, 1 ≤ a ≤ q}. By the usual argument involving the ranks of linear equations whose coefficients involve parameters, it follows that dim H(Ep ) is locally constant on Gp (E) only on the open set where it reaches its minimum. Thus, r is constant on G∗p (E), as we wished to show. Proposition 1.14. Let I ⊂ Ω∗ (M ) be a differential ideal which contains no nonzero forms of degree 0. If ϕ1 , . . . , ϕcn−1 is a polar sequence for the ordinary integral flag (0) ⊂ E1 ⊂ · · · ⊂ En ⊂ Tz M , then it is also a polar sequence for all nearby integral flags.

Proof. Obvious.

We conclude this section by proving a technical proposition which provides an effective method of computing the numbers ci which are associated to an integral flag. We need the following terminology: If J = (j1 , j2 , . . . , jp ) is a multi-index of degree p taken from the set {1, 2, . . . , n}, then we define sup J to be the largest of the integers {j1 , j2 , . . . , jp }. If J = ∅ is the (unique) multi-index of degree 0, we define sup J = 0. Proposition 1.15. Let I ⊂ Ω+ (M ) be an ideal which contains no non-zero forms of degree 0. Let E ∈ Vn (I) be based at z ∈ M . Let ω1 , . . . , ωn , π 1 , s . . . , π (where s = dim M − n) be a coframing on a z-neighborhood so that E = {v ∈ Tz M | π a (v) = 0 for all a}. For each p ≤ n, define Ep = {v ∈ E | ωk (v) = 0 for all k > p}. Let {ϕ1 , ϕ2 , . . . , ϕr } be a set of forms which generate I algebraically where ϕρ has degree dρ + 1. Then, for each ρ, there exists an expansion ρ ϕρ = πJ ∧ ω J + ϕ ˜ρ |J|=dρ

where the 1-forms πJρ are linear combinations of the π’s and the terms in ϕ˜ρ are either of degree 2 or more in the π’s or else vanish at z. Moreover, we have the formula H(Ep ) = {v ∈ Tz M | πJρ (v) = 0 for all ρ and sup J ≤ p}. In particular, for the integral flag (0)z ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ Tz M of I, cp is the number of linearly independent 1-forms in the set {πJρ |z | sup J ≤ p}. Proof. The existence of the expansion cited for ϕρ is an elementary exercise in exterior algebra using the fact that E is an integral element of I. The “remainder term” ϕ˜ρ has the property that ϕ˜ρ (v, e1 , e2 , . . . , edρ ) = 0 for all v ∈ Tz M and all {e1 , e2 , . . . , edρ } ⊂ E. If e1 , e2 , . . . , en is the basis of E which is dual to the coframing ω1 , ω2 , . . . , ωn and K = (k1 , k2, . . . , kdρ ) is a multi-index with deg K = dρ , we have ρ (v). The stated formulas for H(Ep ) and cp follow the formula ϕρ (v ∧ eK ) = πK immediately. To see the utility of Proposition 1.15, consider Example 1.12. Here, I is generated by two forms ϕ1 = ϑ and ϕ2 = −dϑ, of degrees 2 and 3 respectively. If E ∈

70


G3 (T R6 , Ω) is any integral element, then the annihilator of E is spanned by 1forms πi = dui − pij dxj for some numbers pij . It is clear that we have expansions of the form ϕ1 = ϑ = π1 ∧ dx1 + π2 ∧ dx2 + π3 ∧ dx3 + ϕ ˜1 ϕ2 = −dϑ = π3 ∧ dx1 ∧ dx2 + π2 ∧ dx3 ∧ dx1 + π1 ∧ dx2 ∧ dx3 + ϕ ˜2 , on a neighborhood of the base point of E. By Proposition 1.15, it follows that the annihilator of H(E1 ) is spanned by {π1 } and the annihilator of H(E2 ) is spanned by {π1 , π2, π3 }. Thus, we must have c1 = 1 and c2 = 3, as we computed before.

§2. The Cartan–K¨ ahler Theorem. In this section, we prove the Cartan–K¨ ahler theorem, which is the fundamental existence result for integral manifolds of a real-analytic differential system. This theorem is a coordinate-free, geometric generalization of the classical Cauchy– Kowalevski theorem, which we now state. We shall adopt the index ranges 1 ≤ i, j ≤ n and 1 ≤ a, b ≤ s. Theorem 2.1 (Cauchy–Kowalevski). Let y be a coordinate on R, let x = (xi ) be coordinates on Rn , let z = (z a ) be coordinates on Rs , and let p = (pai ) be coordinates on Rns . Let D ⊂ Rn × R × Rs × Rns be an open domain, and let G : D → Rs be a real analytic mapping. Let D0 ⊂ Rn be an open domain and let f : D0 → Rs be a real analytic mapping so that the “1-graph” (11)

Γf = {(x, y0 , f(x), Df(x)) | x ∈ D0 }

lies in D for some constant y0 . (Here, Df(x) ∈ Rns , the Jacobian of f, is described by the condition that pai (Df(x)) = ∂f a (x)/∂xi .) Then there exists an open neighborhood D1 ⊂ D0 × R of D0 × {y0 } and a real analytic mapping F : D1 → Rs which satisfies the P.D.E. with initial condition ∂F/∂y = G(x, y, F, ∂F/∂x) (12)

F (x, y0 ) = f(x) for all x ∈ D0 .

Moreover, F is unique in the sense that any other real-analytic solution of (12) agrees with F on some neighborhood of D0 × {y0 }. We shall not prove the Cauchy–Kowalevski theorem here, but refer the reader to other sources, such as Trêves [1975] or Spivak [1979]. We remark, however, that the assumption of real analyticity is necessary in both the function G (which defines the system of P.D.E.) and the initial condition f. In the smooth category, there are examples where the existence part of the above statement fails and there are other examples where the uniqueness part of the above statement fails. We now turn to the statement of the Cartan–K¨ ahler theorem. If I ⊂ Ω∗ (M ) is a differential ideal, we shall say that an integral manifold of I, V ⊂ M , is a K¨ ahler-regular integral manifold if the tangent space Tv V is a Kähler-regular integral element of I for all v ∈ V . If V is a connected, Kähler-regular integral manifold of I, then we define r(V ) to be r(Tv V ) where v is any element of V .

§2. The Cartan–K¨ ahler Theorem

71

Theorem 2.2 (Cartan–Kähler). Let I ⊂ Ω∗ (M ) be a real analytic differential ideal. Let P ⊂ M be a connected, p-dimensional, real analytic, K¨ ahler-regular integral manifold of I. Suppose that r = r(P ) is a non-negative integer. Let R ⊂ M be a real analytic submanifold of M which is of codimension r, which contains P , and which satisfies the condition that Tx R and H(Tx P ) are transverse in Tx M for all x ∈ P . Then there exists a real analytic integral manifold of I, X, which is connected and (p + 1)-dimensional and which satisfies P ⊂ X ⊂ R. This manifold is unique in the sense that any other real analytic integral manifold of I with these properties agrees with X on an open neighborhood of P . Proof. The theorem is local, so it suffices to prove existence and uniqueness in a neighborhood of a single point x0 ∈ P . Let s = dim M − (r + p + 1). (The following proof holds with the obvious simplifications if any of p, r, or s are zero. For simplicity of notation, we assume that they are all positive.) Our hypothesis implies that the vector space Tx R ∩ H(TxP ) has dimension p + 1 for all x ∈ P . It follows that we may choose a local (real analytic) system of coordinates centered on x0 of the form x1 , . . . , xp , y, u1 , . . . , us , v1 , . . . , vr so that P is given in this neighborhood by the equations y = u = v = 0, R is given in this neighborhood by the equations v = 0, and, for all x ∈ P , the polar space H(Tx P ) is spanned by the vectors {∂/∂xj }1≤j≤p ∪ {∂/∂y} ∪ {∂/∂vρ }1≤ρ≤r . Now, there exists a neighborhood U of Tx0 P in Gp (T M ) so that every E ∈ U with base point z ∈ M has a basis of the form Xi (E) = (∂/∂xi + qi (E)∂/∂y + pσi (E)∂/∂uσ + wiρ (E)∂/∂vρ )|z . The functions x, y, u, v, q, p, w form a coordinate system on U centered on Tx0 P . By the definition of H(Tx0 P ), there exist s real analytic (p + 1)-forms κ1 , . . . , κs in I with the property that H(Tx0 P ) = {v ∈ Tx0 M | κσ (v, ∂/∂x1 , ∂/∂x2 , . . . , ∂/∂xp) = 0 for 1 ≤ σ ≤ s}. In fact, we may even assume that κσ (v, ∂/∂x1 , ∂/∂x2 , . . . , ∂/∂xp ) = duσ (v) for 1 ≤ σ ≤ s and all v ∈ Tx0 M . By the K¨ ahler-regularity of Tx0 P , we may assume, by shrinking U if necessary, that, for all E ∈ Vp (I) ∩ U with base point z ∈ M , we have H(E) = {v ∈ Tz M | κσ (v, X1 (E), X2 (E), . . . , Wp (E)) = 0 for 1 ≤ σ ≤ s}. If we seek v ∈ H(E) of the form v = (a∂/∂y + bσ ∂/∂uσ + cρ ∂/∂vρ )|z , then the s equations κσ (v, X1 (E), X2 (E), . . . , Xp (E)) = 0 are, of course, linear equations for the quantities a, b, and c of the form Aσ (E)a + Bτσ (E)bτ + Cρσ (E)cρ = 0. Again, by hypothesis, when E = Tx0 P these s equations are linearly independent and reduce to the equations bσ = 0. Thus, by shrinking U if necessary, we may

72


assume that the s × s matrix B(E) = (Bτσ (E)) is invertible for all E ∈ U . It follows that there exist unique real analytic functions Gσ on U so that, for each E ∈ U based at z ∈ M , the vector Y (E) = (∂/∂y + Gσ (E)∂/∂uσ )|z satisfies κσ (Y (E), X1 (E), X2 (E), . . . , Xp (E)) = 0. Since the functions x, y, u, v, q, p, w form a coordinate system on U centered on Tx0 P , we may regard the functions Gσ as functions of these variables. We first show that there exists a real analytic submanifold of R of the form v = 0, u = F (x, y) on which the forms κσ vanish. Note that the following vectors would be a basis of the tangent space to such a submanifold at the point z(x, y) = (x, y, F (x, y), 0): Xi (x, y) = (∂/∂xi + ∂i F σ (x, y)∂/∂uσ )|z(x,y) Y (x, y) = (∂/∂y + ∂y F σ (x, y)∂/∂uσ )|z(x,y). It follows that the function F would have to be a solution to the system of P.D.E. given by (13)

∂y F σ = Gσ (x, y, F, 0, 0, ∂xF, 0).

Moreover, in order that the submanifold contain P (which is given by the equations y = u = v = 0), it is necessary that the function F satisfy the initial condition (14)

F (x, 0) = 0.

Conversely, if F satisfies (13) and (14), then the submanifold of R given by v = 0 and u = F (x, y) will both contain P and be an integral of the set of forms {κσ }1≤σ≤s . By the Cauchy–Kowalevski theorem, there exists a unique real analytic solution F of (13) and (14). We let X ⊂ R denote the (unique) submanifold of dimension p + 1 constructed by this method. Replacing the functions u in our coordinate system by the functions u − F (x, y) will not disturb any of our normalizations so far and allows us to suppose, as we shall for the remainder of the proof, that X is described by the equations u = v = 0. We must now show that X is an integral manifold of I. We have already seen that X is the unique connected real analytic submanifold of dimension p + 1 which satisfies P ⊂ X ⊂ R and is an integral of the forms {κσ }1≤σ≤s . We now show that all of the p-forms in I vanish on X. Again using the K¨ ahler-regularity of Tx0 P , let β 1 , . . . , β a be a set of real analytic p-forms in I so that the functions f c (E) = β c (X1 (E), . . . , Xp (E)) for 1 ≤ c ≤ a have linearly independent differentials on U and have the locus Vp (I) ∩ U as their set of common zeros. (We may have to shrink U once more to do this.) Since Tx0 X lies in Vp+1 (I) by construction, Proposition 1.8 shows that Tx0 X is K¨ ahler-ordinary. In fact, the proof of Proposition 1.8 shows that the (p + 1)-forms {β c ∧ dy}1≤c≤a ∪ {κσ }1≤σ≤s have Vp+1 (I) ∩ U + as their set of ordinary common zeros in some neighborhood U + of Tx0 X in Gp+1 (T M ). Thus, in order to show that X is an integral manifold of I, it suffices to show that the forms {β c ∧ dy}1≤c≤a vanish on X.


73

Shrinking U + if necessary, we may suppose that every E + ∈ U + has a basis X1 (E + ), X2 (E + ), . . . , Xp (E + ), Y (E + ) that is dual to the basis of 1-forms dx1 , dx2 , . . . , dxp, dy. If we set B c (E + ) = β c ∧ dy(X1 (E + ), X2 (E + ), . . . , Xp (E + ), Y (E + )) K σ (E + ) = κσ (X1 (E + ), X2 (E + ), . . . , Xp (E + ), Y (E + )), then we have Vp+1 (I) ∩ U + = {E + ∈ U + | B c (E + ) = K σ (E + ) = 0}. Since I is an ideal, the forms β c ∧dxi are also in I and hence vanish on Vp+1 (I)∩U + . Thus, if we set B ci (E + ) = β c ∧ dxi (X1 (E + ), X2 (E + ), . . . , Xp (E + ), Y (E + )), then the functions B ci are in the ideal generated by the functions B c and K σ . It follows that there exist real analytic functions A and L on U + so that b ci σ B ci = Aci b B + Lσ K .

Since K σ (Tz X) = 0 for all z ∈ X by construction, it follows that b B ci (Tz X) = Aci b (Tz X)B (Tz X)

for all z ∈ X. Since I is differentially closed, the forms dβ c are in I. Thus, if we set Dc (E + ) = dβ c (X1 (E + ), X2 (E + ), . . . , Xp (E + ), Y (E + )), there must exist functions G and H on U + so that Dc = Gcb B b + Hσc K σ . Again, since K σ (Tz X) = 0, we must have Dc (Tz X) = Gcb(Tz X)B b (Tz X) for all z ∈ X. Now, if we restrict the forms β c to X, then we have an expansion of the form β c |X = B c (x, y)dx1 ∧ · · · ∧ dxp + (−1)p−i+1 B ci (x, y)dx1 ∧ · · · ∧ dxi−1 ∧ dxi+1 ∧ · · · ∧ dxp ∧ dy, i

where, for z = (x, y, 0, 0) ∈ X, we have set B c (x, y) = B c (Tz X) and B ci (x, y) = B ci (Tz X). We also have the formula c p c ci ∂i B (x, y) dx1 ∧ · · · ∧ dxp ∧ dy dβ |X = (−1) ∂y B (x, y) + i

= D (x, y)dx ∧ · · · ∧ dx ∧ dy, c

1

p

74


where we have written Dc (x, y) = Dc (Tz X) for z as before. Using the formulas Dc (x, y) = Gcb(x, y)B b (x, y) b B ci (x, y) = Aci b (x, y)B (x, y),

we see that the functions B c (x, y) satisfy a linear system of P.D.E. of the form b b ˜c ∂y B c (x, y) = A˜ci b (x, y)∂i B (x, y) + Gb (x, y)B (x, y)

˜ on X. Moreover, since Tz X is an integral element of for some functions A˜ and G I when y = 0, we have the initial conditions B c (x, 0) = 0. By the uniqueness part of the Cauchy–Kowalevski theorem and the fact that all of the functions involved are real-analytic, it follows that the functions B c (x, y) must vanish identically. In turn, this implies that the forms β c vanish on X. Hence X is an integral manifold of I, as we wished to show. Since we have already established uniqueness, we are done. The role of the “restraining manifold” R in the Cartan–K¨ ahler theorem is to convert the “underdetermined” Cauchy problem one would otherwise encounter in extending P to a (p + 1)-dimensional integral to a determined problem. In the coordinate system we introduced in the proof, we could have taken, instead of R, ˜ defined by the which was defined by the equations vρ = 0, the submanifold R, equations vρ = f ρ (x, y) where the functions f ρ are “small” but otherwise arbitrary real analytic functions of the p + 1 variables (x1 , x2 , . . . , xp, y). The construction ˜ of dimension in the above proof would then have lead to an integral manifold X σ σ ρ ρ p + 1 defined locally by equations u = g (x, y) and v = f (x, y). In this sense, the (p + 1)-dimensional extensions of a given p-dimensional, K¨ ahler-regular integral manifold P of I depend on r(P ) functions of p + 1 variables. The Cartan–Kähler theorem has the following extremely useful corollary. (In fact, this corollary is used more often than Theorem 2.2 and is often called the Cartan–K¨ ahler theorem, even in this book.) Corollary 2.3. Let I be an analytic differential ideal on a manifold M . Let E ⊂ Tx M be an ordinary integral element of I. Then there exists an integral manifold of I which passes through x and whose tangent space at x is E. Proof. Assume that the dimension of E is n and let (0)x = E0 ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En = E ⊂ Tx M be an ordinary integral flag. Suppose that, for some p < n, we have found a p-dimensional, regular, real analytic integral manifold Xp of I which passes through x and which satisfies Tx Xp = Ep . Then it is easy to see that there exists a real analytic manifold Rp ⊂ M which contains Xp , is of codimension r(Ep ), and satisfies Tx Rp ∩ H(Ep ) = Ep+1 . Shrinking Xp if necessary, we may assume that Tz Rp is transverse to H(Tz Xp ) for all z ∈ Xp . Applying Theorem 2.2, we see that there exists a real analytic integral manifold of I of dimension (p + 1), Xp+1 , with the property that Tx Xp+1 = Ep+1 . If p + 1 < n, then Ep+1 is a Kähler-regular integral element of I and hence, by shrinking Xp+1 if necessary, we may assume that Xp+1 is a (p + 1)-dimensional, K¨ ahler-regular, real analytic integral manifold of I. If p + 1 = n, then Xp+1 is the desired integral manifold. We conclude this section by explaining some classical terminology regarding the “generality” of the space of ordinary integral manifolds of an analytic differential


75

system I. For simplicity, let us suppose that I is a real analytic differential system on a manifold M and that I contains no non-zero 0-forms. Let (0)z = E0 ⊂ E1 ⊂ · · · ⊂ En ⊂ Tz M be an ordinary integral flag of I. As usual, for 0 ≤ k ≤ n − 1, let ck be the codimension of H(Ek ) in Tz M . For convenience of notation, let us set c−1 = 0 and cn = s = dim M − n. Then we may choose a z-centered coordinate system on some z-neighborhood of the form x1 , x2 , . . . , xn, u1 , u2 , . . . , us so that Ek is spanned by the vectors {∂/∂xj }1≤j≤k and so that, for k < n, H(Ek ) = {v ∈ Tz M | dua (v) = 0 for all a ≤ ck }. For any integer a between 1 and s, let λ(a), the level of a, be the integer k between 0 and n which satisfies ck−1 < a ≤ ck . The number of integers of level k is clearly ck − ck−1 . The (non-negative) number sk = ck − ck−1 is called the k th Cartan character of the given integral flag. Let Ω = dx1 ∧ dx2 ∧ · · ·∧ dxn and let Vn (I, Ω) denote the space of n-dimensional ˜ ∈ Vn (I, Ω), let integral elements of I on which Ω does not vanish. For each E j ˜ ˜ us define Ek = {v ∈ E | dx (v) = 0 for all j > k}. Then for each k, the map E˜ → E˜k is a continuous mapping from Vn (I, Ω) to Vk (I). It follows that there exists a (connected) neighborhood U of En in Vn (I, Ω) with the property that E˜k is K¨ ahler-regular for all k < n and all E˜ ∈ U . By shrinking U if necessary, we may even suppose that the 1-forms {dxj }1≤j≤n and {dua }a>ck are linearly independent ˜k ) for all k < n and all E˜ ∈ U . on H(E We now want to give a description of the collection C of real analytic n-dimensional integral manifolds of I whose tangent spaces all belong to U and which intersect the locus x = 0. By Corollary 2.3, we know that C is non-empty. If X belongs to C, then locally, we may describe X by equations of the form ua = F a (x1 , x2 , . . . , xn ). If the index a has level k, let us define f a to be the function of k variables given by f a (x1 , x2 , . . . , xk ) = F a(x1 , x2 , . . . , xk , 0, 0, . . ., 0). By convention, for level 0 we speak of “functions of 0 variables” as “constants”.) Then the collection {f a }1≤a≤s is a set of s0 constants, s1 functions of 1 variable, s2 functions of 2 variables, . . . , and sn functions of n variables. We now claim that the collection {f a }1≤a≤s characterizes X in the sense that ˜ in C which gives rise to the same collection of functions {f a }1≤a≤s agrees any X with X on a neighborhood of the point (x, u) = (0, F (0)). Moreover, the functions in the collection {f a }1≤a≤s are required to be “small”, but are otherwise arbitrary. It is in this sense that C is parametrized by s0 constants, s1 functions of 1 variable, s2 functions of 2 variables, . . . , and sn functions of n variables. It is common to interpret this as meaning that the local n-dimensional integrals of I depend on s0 constants, s1 functions of 1 variable, s2 functions of 2 variables, . . . , and sn functions of n variables. To demonstrate our claim, let {f a }1≤a≤s be a collection of real analytic functions which are suitably “small” and where f a is a function of the variables x1 , x2 , . . . , xλ(a). For 1 ≤ k ≤ n, define the manifold Rk to be the locus of the equations xk+1 = xk+2 = · · · = xn = 0 and ua = f a (x1 , . . . , xk , 0, . . . , 0) where a ranges over all indices of level greater than or equal to k. The codimension of Rk is n−k+(s−ck−1 ) = r(Ek−1 ). Define X0 to be the point (x, u) = (0, f(0)). By sucessive applications of the Cartan–Kähler theorem, we may construct a unique nested sequence of integral manifolds of I, {Xk }0≤k≤n, which also satisfy the conditions Xk ⊂ Rk . (It is at

76


this stage that we use the assumption of “smallness” in order to guarantee that the necessary transversality conditions hold.) This clearly demonstrates our claim. The following proposition shows that the sequence of Cartan characters has an invariant meaning. Proposition 2.4. Let I ⊂ Ω∗ (M ) be a smooth differential system which contains no 0-forms. Let Z ⊂ Vn (I) be a component of the space of ordinary integral elements. Then the sequence of Cartan characters (s0 , s1 , . . . , sn ) is the same for all ordinary integral flags (0)z ⊂ E1 ⊂ · · · ⊂ En with En ∈ Z. Proof. This follows immediately from Proposition 1.13 and the definitions of the sk , namely: s0 = c0 , sk = ck − ck−1 for 1 ≤ k < n, and sn = s − cn−1 . Usually, the component Z must be understood from context when such statements as “The system I has Cartan characters (s0 , s1 , . . . , sn ).” are made. In fact, in most cases of interest, the space of ordinary, n-dimensional integral elements of I has only one component anyway. æ

§3. Examples. In this section, we give some applications of the Cartan–K¨ ahler theorem. Some of the examples are included merely to demonstrate techniques for calculating the quantities one must calculate in order to apply the Cartan–K¨ ahler theorem, while others are more substantial. The most important example in this section is the application of the Cartan–K¨ ahler theorem to the problem of isometric embedding (see Example 3.8). Example 3.1 (The Frobenius theorem). Let M be a manifold of dimension m = n + s and let I be a differential system which is generated algebraically in degree 1 by a Pfaffian system I ⊂ T ∗ M of rank s. Then at each x ∈ M , there is a unique integral element of dimension n, namely Ix⊥ ∈ Tx M . In fact, every integral element of I based at x must be a subspace of Ix⊥ , since H((0)x ) = Ix⊥ . Thus, if (0)x ⊂ E1 ⊂ · · · ⊂ En = Ix⊥ is an integral flag, then we have H(Ep ) = Ix⊥ for all 0 ≤ p ≤ n. Thus cp = s for all p. It follows by Theorem 1.11 that Vn (I) must have codimension at least ns in Gn (T M ). On the other hand, since there is a unique integral element of I at each point of M , it follows that Vn (I) is a smooth manifold of dimension n+s while Gn (T M ) has dimension n+s+ns. Thus, Vn (I) is a smooth submanifold of codimension ns in Gn (T M ). By Theorem 1.11, it follows that all of the elements of Vn (I) are ordinary. If we now assume that I is real analytic, then the Cartan–K¨ ahler theorem applies (in the form of Corollary 2.3) to show that there exists an n-dimensional integral manifold of I passing through each point of M . The characters are s0 = s and sp = 0 for all p > 0. Thus, according to our discussion at the end of Section 2, the local integral manifolds of I of dimension n depend on s constants. This is in accordance with the usual theory of foliations. The assumption of analyticity is, of course, not necessary. We have already proved the Frobenius theorem in the smooth category in Chapter II. The reader might find it helpful to compare this proof with the proof given there.

§3. Examples

77

Example 3.2 (Orthogonal coordinates). Let g be a Riemannian metric on a manifold N of dimension n. We wish to know when there exist local coordinates x1 , x2 , . . . , xn on N so that the metric takes the diagonal form g = g11 (dx1 )2 + g22 (dx2 )2 + · · · + gnn (dxn )2 . Equivalently, we wish the coordinate vector fields {∂/∂xi }1≤i≤n to be orthogonal with respect to the metric. Such a local coordinate system is said to be orthogonal. If x1 , x2 , . . . , xn is such an orthogonal coordinate system, then the set of 1-forms √ ηi = gii dxi is a local orthonormal coframing of N . These forms clearly satisfy the equations ηi ∧ dηi = 0. Conversely, if η1 , η2 , . . . , ηn is a local orthonormal coframing on M which satisfies the equations ηi ∧ dηi = 0, then the Frobenius theorem implies that there exist local functions x1 , x2 , . . . , xn on N so that ηi = fi dxi (no summation) for some non-zero functions fi . It follows that the functions x1 , x2 , . . . , xn form a local orthogonal coordinate system on N . Thus, our problem is essentially equivalent to the problem of finding local orthonormal coframes η1 , η2, . . . , ηn which satisfy the equations ηi ∧ dηi = 0. Let F → N denote the bundle of orthonormal coframes for the metric g on N . Thus, for each x ∈ N , the fiber Fx consists of the set of all orthonormal coframings of the tangent space Tx N . The bundle F has a canonical coframing ωi , ωij = −ωji ´ Cartan: which satisfies the structure equations of E. dωi = −

ωij ∧ ωj

j

dωij = −

ωik ∧ ωkj +

k

1 2

Rijkl ωk ∧ ωl .

k,l

The forms ωi have the “reproducing property” described as follows: If η = (η1 , η2 , . . . , ηn ) is any local orthonormal coframing defined on an open set U ⊂ M , then η may be regarded as a local section η : U → F of F . Then the formula η ∗ (ωi ) = ηi holds. We set Ω = ω1 ∧ ω2 ∧ · · ·∧ ωn . Notice that Ω does not vanish on the submanifold η(U ) ⊂ F since η ∗ (Ω) = η1 ∧ η2 ∧ · · · ∧ ηn = 0. Conversely, it is clear that any n-dimensional submanifold X ⊂ F on which Ω does not vanish is locally of the form η(U ) for some section η. Let I denote the differential system on F generated by the n 3-forms Θi = ωi ∧ dωi . If X ⊂ F is an integral of the system I on which Ω does not vanish, then clearly X is locally of the form η(U ) where η is a local orthonormal coframing satisfying our desired equations ηi ∧ dηi = 0. Conversely, a local orthonormal coframing η satisfying ηi ∧ dηi = 0 has the property that η(U ) is an n-dimensional integral manifold of I on which Ω does not vanish. We proceed to analyse the n-dimensional integral manifolds of I on which Ω does not vanish. Note that I is generated algebraically by the 3-forms Θi and the 4-forms Ψi = dΘi . Let E ∈ Vn (I, Ω) be based at f ∈ F. When we restrict the forms ωi , ωij to E, the forms ωi remain linearly independent, and we have relations of the form Θi = −( ωij ∧ ωj ) ∧ ωi = 0. j

78


It follows that there exist 1-forms λi = Collecting terms, we have the equation

j

Lij ωj so that ( j ωij ∧ ωj ) = λi ∧ ωi .

(ωij + Lij ωi − Lji ωj ) ∧ ωj = 0.

j

Since the forms ϕij = (ωij + Lij ωi − Lji ωj ) are skew-symmetric in their indices, it follows that the above equations can only hold if we have ωij + Lij ωi − Lji ωj = 0. Conversely, we claim that if {Lij }i =j is any set of n2 − n numbers, then the n-plane E ⊂ Tf F annihilated by the 1-forms ϕij = (ωij + Lij ωi − Lji ωj ) is an integral element of I on which Ω does not vanish. To see this, note that for such E, we have the identity −dωi = ( j ωij ∧ ωj ) = λi ∧ ωi . It immediately follows that Θi = ωi ∧ dωi and Ψi = dωi ∧ dωi must vanish on E. It follows that the space of integral elements of I which are based at a point of F is naturally a smooth manifold of dimension n2 − n. Moreover, the space Vn (I, Ω) is a smooth manifold of dimension dim F + (n2 − n). Thus, the codimension of n Vn (I, Ω) in Gn (F ) is (n − 2) 2 . When n = 2, we are looking for integrals of dimension 2. However, I has no non-zero forms of degree less than 3. It follows that any surface in F is an integral of I. 1 2 From now on, we assume that n ≥ 3. Since dim nF = 2 (n + n), if (0)f ⊂ E1 ⊂ · · · ⊂ En is an integral flag, it follows that cp ≤ 2 for all 0 ≤ p ≤ n. However, since I contains no non-zero forms of degree less than 3, it follows that c0 = c1 = 0. Moreover, since I contains only n 3-forms, it follows that c2 ≤ n. Thus, we have the inequality n c0 + c1 + c2 + · · · + cn−1 ≤ n + (n − 3) 2 It follows, by Theorem 1.11, that for n ≥ 4, none of the elements of Vn (I, Ω) are ordinary. Thus, the Cartan–K¨ ahler theorem cannot be directly applied in the case where n ≥ 4. This is to be expected since a Riemannian metric in n variables has n2 “off diagonal” components in a general coordinate system and a choice of coordinates depends on only n functions of n variables. Thus, if n < n2 , (which holds when n ≥ 4) we do not expect to be able to diagonalize the “generic” Riemannian metric in n variables by a change of coordinates. Let us now specialize to the case n = 3. By Theorem 1.11 and the calculation above, an integral element E ∈ V3 (I, Ω) is an ordinary integral element if and only if it contains a 2-dimensional integral element E2 whose polar space has codimension 3. Now a basis for the 3-forms in I can be taken to be {ω23 ∧ ω2 ∧ ω3 , ω31 ∧ ω3 ∧ ω1 , ω12 ∧ ω1 ∧ ω2 }. If E ∈ V3 (I, Ω) is given, then let v1 , v2 ∈ E be two vectors which span a 2-plane E2 on which none of the 2-forms {ω2 ∧ ω3 , ω3 ∧ ω1 , ω1 ∧ ω2 } vanish. Then it immediately follows that the polar equations of E2 have rank 3. Thus E is ordinary. This yields the following theorem. Theorem 3.3 (Cartan). Let (M 3 , g) be a real analytic Riemannian metric. Let S be a real analytic surface in M and let η : S → F be a real analytic coframing along S so that none of the 2-forms {η2 ∧ η3 , η3 ∧ η1 , η1 ∧ η2 } vanishes when restricted

§3. Examples

79

to S. Then there is an open neighborhood U of S in M and a unique real analytic extension of η to U so that the equations ηi ∧ dηi = 0 hold. Proof. The surface η(S) ⊂ F is a K¨ ahler-regular integral of I by the above calculation. The rest follows from the Cartan–K¨ ahler theorem. Corollary 3.4. Let (M 3 , g) be a real analytic Riemannian metric. Then every point of M lies in a neighborhood on which there exists a real analytic orthogonal coordinate system for g. Remark. Theorem 3.3 has been proved in the smooth category by DeTurck and Yang [1984]. Their proof relies on the theory of the characteristic variety of an exterior differential system, and in that context, this example will be revisited in Chapter V. For example, the conditions on the ηi ∧ ηj in the above theorem say exactly that S is non-characteristic. Example 3.5 (Special Lagrangian geometry). This example is due to Harvey and Lawson [1982]. Let M = Cn with complex coordinates z1 , z2 , . . . , zn . Let I be the ideal generated by the 2-form Φ and the n-form Ψ where √ Φ = ( −1/2)(dz1 ∧ d¯ z1 + dz2 ∧ d¯ z2 + · · · + dzn ∧ d¯ zn ) and

Ψ = Re(dz1 ∧ dz2 ∧ · · · ∧ dzn ) z1 ∧ d¯ z2 ∧ · · · ∧ d¯ zn ). = 12 (dz1 ∧ dz2 ∧ · · · ∧ dzn + d¯

Note that I is invariant under the group of motions of Cn generated by the translations and the rotations by elements of SU (n). We want to examine the set Vp (I) for all p. First assume that E ∈ Vp (I) where p is less than n. Let e1 , e2 , . . . , ep be an orthonormal basis for E, where we use the n n standard inner √ product on C . Since, for any √ two vectors v, w ∈ C , the formula Φ(v, w) = −1 v, w holds, it follows that −1 ej , ek = 0 for all j and k. Thus, the vectors e1 , e2 , . . . , ep are Hermitian orthogonal as well as Euclidean orthogonal. Since p < n, it follows that by applying a rotation from SU (n), we may assume k that ek = ∂/∂x where we define the usual real coordinates on Cn by the equation √ k k z = x + −1 yk . It follows that the group of motions of Cn which preserve I acts transitively on the space Vp (I) for all p < n. In particular, the polar spaces of all of the elements of Vp (I) have the same dimension. Since I contains no non-zero 0-forms, Proposition 1.10 now shows that every integral element of I of dimension less than n is K¨ ahler-regular. Thus, every integral element of I of dimension n is ordinary. For each p < n, let Ep be spanned by the vectors {∂/∂xk }k≤p. Then it is easy to compute that, for p < n − 1, H(Ep) = {v ∈ Tz Cn | dyk (v) = 0 for all k ≤ p}. On the other hand, we have H(En−1) = {v ∈ Tz Cn | dyk (v) = 0 for all k ≤ n − 1 and dxn (v) = 0}. Thus, cp = p for all p < n − 1 and cn−1 = n. In particular, note that there are no integral elements of dimension n + 1 or greater, and each E ∈ Vn (I) is the

80


polar space of any of its (n − 1)-dimensional subspaces. It follows that the group of motions of Cn which preserve I acts transitively on Vn (I) as well. Using the technique of calibrations, Harvey and Lawson show that any n-dimensional integral manifold N n ⊂ Cn of I is absolutely area minimizing with respect to compact variations. They call such manifolds special Lagrangian. We may now combine our discussion of the integral elements of I with the Cartan–K¨ ahler theorem to prove one of their results: Theorem 3.6. Every (n − 1)-dimensional real analytic submanifold P ⊂ Cn on which Φ vanishes lies in a unique real analytic n-dimensional integral manifold of I. Remark. Because of the area minimizing property of the n-dimensional integrals of I, it follows that every integral of I is real analytic. Thus, if P ⊂ Cn−1 is an integral of Φ which is not real analytic, then there may be no extension of P to an n-dimensional integral of I in Cn . As an example of such a P , consider the submanifold defined by the equations z n = y1 = y2 = · · · = yn−2 = yn−1 − f(xn−1 ) = 0 where f is a smooth function of xn−1 which is not real analytic. This shows that the assumption of real analyticity in the Cartan–K¨ ahler theorem cannot be omitted in general. Example 3.7 (An equation with degenerate symbol). This example is a generalization of Example 1.12. Let a vector field V be given in R3 . Let λ be a fixed constant. We wish to determine whether there exists a vector field U in R3 which satisfies the system of 3 equations curl U + λU = V. Note that even though this is three equations for three unknowns, this set of equations cannot be put in Cauchy–Kowalevski form. In fact, computing the divergence of both sides of the given equation, we see that U must satisfy a fourth equation λ div U = div V. Of course, if λ = 0, then a necessary and sufficient condition for the existence of such a vector field U is that div V = 0. If λ = 0, then the situation is more subtle. We shall set up a differential system whose 3-dimensional integrals correspond to the solutions of our problem. Let R6 be given coordinates x1 , x2 , x3, u1 , u2 , u3 . We regard x1 , x2 , x3 as coordinates on R3 . If the components of the vector field V are (v1 , v2 , v3 ), let us define the forms α = u1 dx1 + u2 dx2 + u3 dx3 β = u1 dx2 ∧ dx3 + u2 dx3 ∧ dx1 + u3 dx1 ∧ dx2 γ = v1 dx2 ∧ dx3 + v2 dx3 ∧ dx1 + v3 dx1 ∧ dx2 . Now let I be the differential system on R6 generated by the 2-form Θ = dα+λβ− γ. Any 3-dimensional integral manifold of Θ on which the form Ω = dx1 ∧ dx2 ∧ dx3 does not vanish is locally a graph of the form (x, u(x)) where the components of

§3. Examples

81

u(x) determine a vector field U on R3 which satisfies the equation curl U +λU = V . Conversely, any local solution of this P.D.E. gives rise to an integral of I by reversing this process. We now turn to an analysis of the integral elements of I. The pair of forms Θ, dΘ clearly suffice to generate I algebraically. The cases where λ vanishes or does not vanish are markedly different. If λ = 0, then dΘ = −dγ = −(div V )Ω. Thus, if div V = 0, then there cannot be any integral of I on which Ω does not vanish. On the other hand, suppose div V = 0. Then I is generated by Θ alone since then dΘ ≡ 0. If E ∈ V3 (I, Ω), then the annihilator of E is spanned by three 1-forms of the form π i = dui − j Aij dxj for some numbers A. It follows that, at the base point of E, the form Θ can be written in the form Θ = π 1 ∧ dx1 + π 2 ∧ dx2 + π 3 ∧ dx3 . Setting ωi = dxi , we may apply Proposition 1.15 to show that the characters of the associated integral flag satisfy cp = p for all 0 ≤ p ≤ 3. On the other hand, it is clear from this formula for Θ that there exists a 6-parameter family of integral elements of Θ at each point of R6 . By Theorem 1.11, it follows that all of the elements of V3 (I, Ω) are ordinary. The character sequence is given by s0 = 0 and sp = 1 for p = 1, 2, 3. At this point, if we assumed that V were real analytic, we could apply the Cartan–K¨ ahler theorem to show that there exist local solutions to our original problem. However, in this case, an application of the Poincaré lemma will suffice even without the assumption of real analyticity. Now let us turn to the case where λ = 0. If, for any (i, j, k) which is an even permutation of (1, 2, 3), we set πi = dui + 12 [(λuj − vj )dxk − (λuk − vk )dxj ] − λ−1 ∂i vi dxi ; then

Θ = π1 ∧ dx1 + π2 ∧ dx2 + π3 ∧ dx3 λ−1 dΘ = π1 ∧ dx2 ∧ dx3 + π2 ∧ dx3 ∧ dx1 + π3 ∧ dx1 ∧ dx2 .

It follows that any E ∈ V3 (I, Ω) is annihilated by 1-forms of the form ϑi = πi − j pij dxj where, in order to have ΘE = 0, we must have pij = pji and, in order to ahve (dΘ)E = 0, we must have p11 + p22 + p33 = 0. Thus V3 (I, Ω) is a smooth manifold of codimension 4 in G3 (T R6 ). On the other hand, by Proposition 1.15, it follows that c0 = 0, c1 = 1, and c2 = 3 for the integral flag associated to the choice ωi = dxi . Since c0 + c1 + c2 = 4, it follows that all of the elements of V3 (I, Ω) are ordinary. By the Cartan–Kähler theorem, it follows that, if V is real analytic, then there exist (analytic) local solutions to the equation curl U + λU = V for all non-zero constants λ. Example 3.8 (Isometric embedding). We now wish to consider the problem of locally isometrically embedding an n-dimensional manifold M with a given Riemannian metric g into Euclidean space EN where N is some integer yet to be specified. Note that the condition that a map u : M → EN be an isometric embedding is a set of non-linear, first-order partial differential equations for u. Precisely, if

82


x1 , x2 , . . . , xn is a set of local coordinates on M and g = gij dxi ◦ dxj , then the 1 equations that u must satisfy are gij = ∂i u·∂j u. This is 2 n(n+1) for the N unknown components of u. Thus, we do not expect to have any solution if N < 12 n(n + 1). It is the contention of the Cartan–Janet isometric embedding theorem (which we prove below) that such a local isometric embedding is possible if the metric g is real analytic and N = 12 n(n + 1). Note that even though the isometric embedding system is determined if N = 12 n(n + 1), it cannot be put in Cauchy–Kowalevski form for n > 1. We begin by writing down the structure equations for g. Since our results will be local, we may as well assume that we can choose an orthonormal coframing η1 , η2 , . . . , ηn on M so that the equation g = (η1 )2 + (η2 )2 + · · · + (ηn )2 holds. By the fundamental lemma of Riemannian geometry, there exist unique 1-forms on M , ´ Cartan hold: ηij = −ηji , so that the first structure equations of E. dηi = −

ηij ∧ ηj .

j

´ Cartan also hold: The second structure equations of E. dηij + −

ηik ∧ ηkj +

1 2

k

Rijkl ηk ∧ ηl .

k,l

Here, the functions Rijkl are the components of the Riemann curvature tensor and satisfy the usual symmetries Rijkl = −Rjikl = −Rijlk Rijkl + Riklj + Riljk = 0. Let Fn (EN ) denote the bundle over EN whose elements consist of the (n + 1)tuples (x; e1 , e2 , . . . , en ) where x ∈ EN and e1 , e2 , . . . , en are an orthonormal set of vectors in EN . Note that Fn (EN ) is diffeomorphic to EN ×SO(N )/SO(N −n). For several reasons, it is more convenient to work on Fn (EN ) than on the full frame bundle of EN . We shall adopt the index ranges 1 ≤ i, j, k, l ≤ n < a, b, c ≤ N . Let U ⊂ Fn (EN ) be an open set on which there exist real analytic vector-valued functions ea : U → EN with the property that for all f = (x; e1 , e2 , . . . , en ) ∈ U, the vectors e1 , e2 , . . . , en , en+1 (f), . . . , eN (f) form an orthonormal basis of EN . We also regard the components of f as giving vector-valued functions x, ei : U → EN . It follows that we may define a set of 1-forms on U by the formulae ωi = ei · dx ωa = ea · dx ωij = ei · dej = −ωji ωia = ei · dea = −ωai = −ea · dei ωab = ea · deb = −ωba .

§3. Examples

83

Of these forms, the set {ωi } ∪ {ωa } ∪ {ωij }i<j ∪ {ωai } forms a coframing of U. We shall have need of the following structure equations dωi = −

ωij ∧ ωj −

j

dωa = −

ωaj ∧ ωj −

j

dωij = −

ωib ∧ ωb

b

ωab ∧ ωb

b

ωik ∧ ωkj −

k

ωib ∧ ωbj .

b

Now, on M × U, consider the differential system I− generated by the 1-forms {ωi − ηi }i≤n ∪ {ωa}a>n . Let Ω = ω1 ∧ ω2 ∧ · · · ∧ ωn . Proposition 3.9. Any n-dimensional integral of I− on which Ω does not vanish is locally the graph of a function f : M → U with the property that the composition x ◦ f : M → EN is a local isometric embedding. Conversely, every local isometric embedding u : M → EN arises in a unique way from this construction. Proof. First suppose that we have an isometric embedding u : M → EN . Let {Ei } be the orthonormal frame field on M which is dual to the coframing {ηi }. For each z ∈ M , define f(z) = (u(z); du(E1(z)), . . . , du(En(z))). Now consider the graph Γu = {(z, f(z)) | z ∈ M } ⊂ M × U. Clearly, Γu is an integral of I− if and only if f satisfies f ∗ (ωi ) = ηi and f ∗ (ωa ) = 0. However, since ea (f(z)) is normal to the vectors du(Ei (z)) by construction, we have f ∗ (ωa ) = (ea ◦ f) · d(x ◦ f) = (ea ◦ f) · du = 0. Also, since u is an isometric embedding, we have, for all v ∈ Tz M , f ∗ (ωi )(v) = (ei ◦ f) · du(v) = du(Ei (z)) · du(v) = Ei (z) · v = ηi (v). Note also that, on Γu , we have Ω = ω1 ∧ ω2 ∧ · · · ∧ ωn = η1 ∧ η2 ∧ · · · ∧ ηn . Since the latter form is non-zero when projected onto the factor M , it follows that Ω is non-zero on Γu . Now suppose that X ⊂ M × U is an n-dimensional integral manifold of I− on which Ω does not vanish. Then since ωi = ηi on X, it follows that the form η1 ∧ η2 ∧ · · · ∧ ηn also does not vanish on X. It follows that the projection X → M onto the first factor is a local diffeomorphism. Thus we may regard X locally as the graph of a function f:M → U. Now let u = x ◦ f. We claim that u:M → EN is an isometry and moreover that ei ◦ f = du(Ei ). This will establish both parts of the proposition. To see these claims, note that we have f ∗ (ωi ) = ηi and f ∗ (ωa ) = 0. Since f ∗ (ωa ) = f ∗ (ea · dx) = (ea ◦ f) · du = 0, it follows that (ea ◦ f)(z) is normal to du(Ei (z)) for all i and a and z ∈ M . Thus, the vectors du(Ei (z)) are linear combinations of the vectors {(ej ◦ f)(z)}j≤n . On the other hand, for any v ∈ Tz M , we have Ei (z) · v = ηi (v) = f ∗ (ωi )(v) = f ∗ (ei · dx)(v) = (ei ◦ f)(z) · du(v). Using the fact that {Ei (z)} is an orthonormal basis for Tz M and that {(ei ◦ f)(z)} is an orthonormal basis for du(Tz M ), we see that du must be an isometry and that ei ◦ f = du(Ei ), as claimed. We are now going to show that any integral of I− on which Ω does not vanish is actually an integral of a larger system I (defined below). Suppose that X is such an n-dimensional integral. Then let us compute, on X, 0 = d(ωi − ηi ) = −

j

(ωij − ηij ) ∧ ηj .

84


Since the forms ηi are linearly independent on X and since the forms ϑij = (ωij −ηij ) are skew-symmetric in their lower indices, this implies that the forms ϑij must vanish on X. The geometric meaning of this fact is that the Levi–Civita connection of a Riemannian metric is the same as the connection induced by any isometric embedding into Euclidean space. Let us now consider the differential system I on M × U which is generated by the set of 1-forms {ωi − ηi }i≤n ∪ {ωa}n p} form an ordinary flag, and so that H(Ep ) = E ⊕ {w ∈ Q | yσ (w) = 0 for all σ > s − cp } for p < n. Here we are using the notations from the proof of Proposition 1.10 in Chapter III, and we recall our convention that cn = s and c−1 = 0. By definition, there exist r forms ϕ1 , ϕ2 , . . . , ϕr in I so that the forms ϕρ for 1 ≤ ρ ≤ r generate IE algebraically. Here, ϕρ refers to the construction given by (17). These forms have expansions ρ ϕρ = fσJ dyσ ∧ dxJ σ,|J|=pρ ρ for 1 ≤ ρ ≤ r where ϕρ has degree pρ + 1, and the fσJ are some constants. Clearly, IE is real analytic and moreover, as a differential form on E ⊕ Q each ϕρ is closed. By Theorem 4.3, the Cartan–K¨ ahler theorem applies. We now refer to Chapter III. Recall that there we defined the level of an index σ in the range 1 ≤ σ ≤ s to be the integer k (in the range 0 ≤ k ≤ n) so that s − ck < σ ≤ s − ck−1 . We also recall from these that the characters s0 , s2 , . . . , sn of I in a neighborhood of E are defined by

sk = number of σ’s of level k. By the discussion following the proof of the Cartan–K¨ ahler theorem, the real analytic integral manifolds of (IE , ΩE ) are given in a neighborhood of x = 0 by equations of the form yσ = F σ (x1 , x2 , . . . , xn) where the F σ are real analytic and moreover are uniquely specified by knowing the following data: the s0 constants the s1 functions the s2 functions .. . the sn functions

f σ = F σ (0, 0, . . ., 0) f (x1 ) = F σ (x1 , 0, . . . , 0) σ 1 f (x , x2 ) = F σ (x1 , x2 , 0, . . . , 0) .. . σ

f σ (x1 , x2, . . . , xn ) = F σ (x1 , x2 , . . . , xn )

when σ has level 0 when σ has level 1 when σ has level 2 .. . when σ has level n.

Note that due to the fact that the forms ϕρ have constant coefficients, it follows that if yσ = F σ (x1 , x2 , . . . , xn ) is a real analytic solution in a neighborhood of x = 0 and we let Fkσ be the homogeneous term of degree k in the power series expansion of F σ , then yσ = Fkσ is also an integral manifold of IE . It follows also that if the f σ are each chosen to be homogeneous polynomials of degree k in the appropriate variables, then the corresponding F σ will also be homogeneous polynomials of degree k. If we regard the collection Fk = (Fkσ ) as a Q-valued polynomial on E of degree k, then we see that the subspace S k ⊂ Q ⊗ S k (E ∗ ) consisting of those polynomial maps of degree k whose graphs in E ⊕ Q are integral manifolds of IE is a vector space whose dimension is given, for k > 0, by the formula k+1 k+2 k+n−1 + s3 + · · · + sn . dim S k = s1 + s2 1 2 n−1

110

IV. Linear Differential Systems

Note that S 1 = AE by definition. Moreover, we plainly have that S k = (AE )(k−1) for all k ≥ 1. We now claim that the characters s1 , s2 , . . . , sn of I in a neighborhood of E and s1 , s2 , . . . , sn of the tableau AE are related by sk = sk + · · · + sn .

(∗)

Once we have established then we are done, since it follows that the dimension of (AE )(1) ∼ = S 2 is given by s1 + s2

3 4 n+1 + s3 + · · · + sn 1 2 n−1

= (s1 + · · · + sn ) + 2(s2 + · · · + sn ) + · · · + nsn = s1 + 2s2 + · · · + nsn by (∗). To establish (∗), we have from Definition 3.5 that dim(AE )k = sk+1 + · · · + sn , and we also have from the definition that (AE )k consists of the Q-valued linear functions on E that lie in AE ⊂ Q⊗E ∗ and that do not depend on x1 , . . . , xk . Thus, (AE )k is isomorphic to the space of linear integral manifolds yσ = F σ (xk+1 , . . . , xn ) of (IE , ΩE ) as described above and which do not depend on x1 , . . . , xk . By the count of the number of such solutions there are (starting from the top) (n − k)sn

− dimensions worth coming from an arbitrary linear function f σ (xk+1 , . . . , xn ) where σ has level n

(n − k − 1)sn−1 − dimensions worth coming from an arbitrary linear function f σ (xk+1 , . . . , xn−1) where σ has level n − 1 .. . − dimensions worth coming from an arbitrary linear function f σ (xk+1 ) where σ has level k + 1.

sk+1

Thus dim(AE )k = sk+1 + 2sk+2 + · · · + (n − k)sn . This gives the equations sk+1 + · · · + sn = sk+1 + 2sk+2 + · · · + (n − k)sn which may then be solved to give (∗).

for 0 ≤ k ≤ n − 1,

§5. Linear Pfaffian Systems

111

§5. Linear Pfaffian Systems. The general theory takes a concrete and simple form for Pfaffian systems. Definition 5.1. A Pfaffian system is an exterior differential system with independence condition (I, Ω) such that I is generated as an exterior differential system in degrees zero and one. Locally I is algebraically generated by a set f1 , . . . , fr of functions and θ1 , . . . , θs0 of 1-forms together with the exterior derivatives df1 , . . . , dfr and dθ1 , . . . , dθs0 . Equating the functions fi to zero effectively means restricting to submanifolds, and we shall not carry this step along explicitly in our theoretical developments. However, in practice it is obviously important; for example, in imposing integrability conditions during the process of prolongation (cf. §6 below). Thus, unless mentioned to the contrary, we shall assume that I is generated as a differential ideal by the sections of a sub-bundle I ⊂ T ∗ M .4 Moreover, as explained following Definition 1.1 above, we shall assume that the independence condition corresponds to a sub-bundle J ⊂ T ∗ M with I ⊂ J. Denoting by {J} ⊂ Ω∗ (M ) the algebraic ideal generated by the C ∞ sections of J, we shall prove the following Proposition 5.2. The necessary and sufficient condition that (I, Ω) be linear is that dI ≡ 0 mod {J}.

(46)

Proof. We will use the proof as an opportunity to derive the local structure equations of a Pfaffian system (I, Ω). Choose a set of 1-forms θ 1 , . . . , θ s0 ; ω 1 , . . . , ω n ; π 1 , . . . , π t that is adapted to the filtration I ⊂ J ⊂ T ∗M and that gives a local coframing on M . Throughout this section we shall use the ranges of indices 1 ≤ a, b ≤ s0 1 ≤ i, j ≤ n 1 ≤ ε, δ ≤ t. The above local coframing is defined up to invertible linear substitutions θã = Aab θb (47)

ω ˜ i = Bji ωj + Bai θa π ˜ ε = Cδε π δ + Ciε ωi + Caε θa ,

4 Although it is somewhat cumbersome, we shall use s to denote the rank of I; this s is the 0 0 first of the Cartan characters s0 , s1 , . . . , sn .

112


reflecting the filtration I ⊂ J ⊂ T ∗ M . The behavior of the 2-forms dθa is fundamental to the differential system. In terms of them we will determine the conditions that (I, Ω) be linear. For this we write (48)

1 1 dθa ≡ Aaεi π ε ∧ ωi + caij ωi ∧ ωj + eaεδ π ε ∧ π δ 2 2

mod {I}

where we can suppose that caij + caji = 0 = eaεδ + eaδε . Here we recall our notation that {I} is the algebraic ideal generated by the θa ’s. Integral elements of (I, Ω) are defined by θa = 0 together with linear equations π ε − pεi ωi = 0

(49) where by (48) (50)

(Aaεi pεj − Aaεj pεi ) + caij + eaεδ (pεi pδj − pεj pδi ) = 0.

These equations are linear in pεi if, and only if, eaεδ = 0. This is equivalent to (51)

1 dθa ≡ Aaεi π ε ∧ ωi + caij ωi ∧ ωj 2

mod {I},

which is also clearly the condition that (I, Ω) be linearly generated. It is also clear that (51) is just (46) written out in terms of bases. The proof shows that a Pfaffian system is linear if, and only if, it is linearly generated. Definition 5.3. We shall say that the Pfaffian system (I, Ω) is linear if either of the equivalent conditions (46) or (51) is satisfied. We want to comment on the equation (51). Assume that (I, Ω) is linear and write (52)

dθa ≡ πia ∧ ωi

mod {I}

where the πia are 1-forms. It is clear from (47) that (53)

the πia are well-defined as sections of T ∗ M/J, and under a change of coframe (47) the πia transform like the components, relative to our chosen coframe, of a section of I ∗ ⊗ J/I.


113

The relation between (51) and (52) is πia ≡ Aaεi π ε

(54)

mod {θa , ωi },

and the most general 1-forms πia satisfying (52) are given by πia = Aaεi π ε + caji ωj + paji ωj where the Aaεi and caji are as above and the paij are free subject to paij = paji . In intrinsic terms, for linear Pfaffian systems the exterior derivative induces a bundle mapping δ : I → (T ∗ M/J) ⊗ J/I

(55) given locally by (56)

δ(θa ) = Aaεi π ε ⊗ ωi

where the ωi are viewed as sections of J/I and the π ε are viewed as sections of T ∗ M/J.5 Example 5.4. Referring to Example 1.3, we consider a partial differential equation of second order F (xi , z, ∂z/∂xi , ∂ 2 z/∂xi ∂xj ) = 0. This is equivalent to the Pfaffian differential system ⎧ i ⎪ ⎨ F (x , z, pi , pij ) = 0 (57) θ = dz − pi dxi = 0 ⎪ ⎩ θi = dpi − pij dxj = 0 with independence condition dx1 ∧· · ·∧dxn = 0 in the space of variables xi , z, pi , pij = pji . The exterior derivatives of θ and θi are clearly in the algebraic ideal generated by dxi , θ, θi . Hence the Pfaffian system is linear. (58) Remark. In general, it is true that (i) the contact systems on the jet spaces J k (Rn , Rs0 ) are linear Pfaffian systems, and (ii) the restriction of a linear Pfaffian system to a submanifold is again a linear Pfaffian system (assuming that the independence form Ω is non-zero modulo the ideal I on the submanifold). Hence this example is valid for a P.D.E. system of any order. Example 5.5. Referring to Example 1.5, the canonical system (L, Φ) on Gn (T M ) is a linear Pfaffian system. We next want to define the important concept of the tableau bundle associated to a linear Pfaffian system (I, Ω) satisfying suitable constant rank conditions. This will be a sub-bundle A ⊂ I ∗ ⊗ J/I 5 This

flag.

δ¯ is closely related to the maps δ = d mod I encountered in the discussion of the derived

114


given for each x ∈ M by a tableau Ax ⊂ Ix∗ ⊗ Jx /Ix as defined in §3 above, and with the property we always assume that dim Ax is locally constant. To define Ax we let Jx⊥ ⊂ Tx M be given as usual by Jx⊥ = {v ∈ Tx M : η(v) = 0 for all η ∈ J} = {v ∈ Tx M : θa (v) = ωi (v) = 0}. Then, referring to (53) above, the quantities πia (v) ∈ Ix∗ ⊗ Jx /Ix ,

v ∈ Jx⊥

are well defined, and we set (59)

Ax = {πia (v) : v ∈ Jx⊥ }.

More precisely, the choice of framing θa for I and ωi for J/I induce bases wα and xi for Ix∗ and Jx /Ix , respectively. Then Ax is spanned by the quantities π(v) = πia (v)wa ⊗ xi for v ∈ Jx⊥ . Definition 5.6. Assuming that dim Ax is locally constant on M , we define the tableau bundle A ⊂ I ∗ ⊗ J/I by the condition that its fibres be given by (59). Remark. We observe from (52) that the mapping Jx⊥ → Ix∗ ⊗ Jx /Ix given by v → πia (v) is injective if, and only if, there are no vectors v ∈ Jx⊥ satisfying v

Ix ⊂ Ix .

In particular, this is the case if there are no Cauchy characteristic vectors for I, and in this situation the tableau Ax has as basis the matrices Aε = Aaεi ,

ε = 1, . . . , t.

In general, these matrices Aε span Ax but may not give a basis. In intrinsic terms, by dualizing (55) with respect to T ∗ M/J and I we have a bundle mapping (60)

π : J ⊥ → I ∗ ⊗ J/I,

and with our constant rank assumption the tableau bundle is the image π(J ⊥ ). The mapping π is given in the above coordinates by π(v) = πia (v)wa ⊗ xi .


115

From (51) it is clear that for linear Pfaffian systems the tableau bundle encodes what we might call the “principal part” of the behavior of the 2-forms dθa mod {I}. Here, principal part refers to the term Aaεi π ε ∧ ωi ; the other term 12 caij ωi ∧ ωj will also be discussed below. For each x ∈ M , the characters si (x) of the tableau Ax are defined; we shall assume that these are locally constant and shall call them the reduced characters of the linear Pfaffian system. In the discussion of examples it will frequently be convenient to let x ∈ M be a typical point, set ⎧ ∗ ⎪ ⎨ W = Ix V ∗ = Jx /Ix ⎪ ⎩ A = Ax and si = si (x), and speak of A ⊂ W ⊗V ∗ as the tableau of (I, Ω) without reference to the particular point x ∈ M . In Definition 3.4 above, we introduced the prolongation of a tableau and in (26) we gave an interpretation of the prolongation. This interpretation may be extended as follows: Proposition 5.7. Assume that (I, Ω) is a linear Pfaffian system and that the set of integral elements Gx(I, Ω) of (I, Ω) lying over x ∈ M is non-empty. Then Gx (I, Ω) is an affine linear space whose associated vector space may be naturally (1) identified with the prolongation Ax of Ax . Proof. We work over a fixed point x ∈ M and omit reference to it. Referring to the proof of Proposition 5.2, the equations of integral elements are

θa = 0 π ε − pεi ωi = 0

where (Aaεi pεj − Aaεj pεi ) + caij = 0. These equations define an affine linear space, and assuming that this is non-empty (a point we shall take up next) the associated vector space is defined by the homogeneous linear equations (61)

Aaεi pεj = Aaεj pεi .

a and see that Given a solution pεj to these equations, then we set Pija = Aaεi pεj = Pji

P = Pija wa ⊗ xi xj ∈ W ⊗ S 2 V ∗ satisfies the relations Baλi Pija = 0 that define A; hence P ∈ A(1) . Conversely, if P = Pija wa ⊗ xi xj ∈ A(1) ,

116


then each ∂P/∂xj ∈ A and so is a linear combination ∂P/∂xj = Aaεi pεj wa ⊗ xi of a spanning set of matrices Aε = Aaεi of A. The condition ∂ 2 P/∂xi ∂xj = ∂ 2 P/∂xj ∂xi is then equivalent to (61). In §3, we have defined the symbol associated to a tableau, and here we have the corresponding Definition 5.8. Let (I, Ω) be a linear Pfaffian system with tableau bundle A ⊂ I ∗ ⊗ J/I. Then the symbol bundle is defined to be B = A⊥ ⊂ I ⊗ (J/I)∗ . As explained above, we shall frequently omit reference to the point x ∈ M and simply refer to B as the symbol of the Pfaffian differential system. Example 5.9. We consider a 1st order P.D.E. system (62)

F λ (xi , ya , ∂ya /∂xi ) = 0.

We write this as the Pfaffian differential system (63)

θa = dya − pai dxi = 0 Ω = dx1 ∧ · · · ∧ dxn = 0

in the submanifold M of (xi , ya , pai ) space defined by the equations (64)

F λ (xi , ya , pai) = 0

(we assume that these define a submanifold). The structure equations of (63) are dθa ≡ πia ∧ ωi

mod (I)

where πia = −dpai |M and ωi = dxi |M . From (64) the πia are subject to the relations ∂F λ a π ≡ 0 mod {J}. ∂pai i It follows from the above discussion that at each point q = (xi , ya , pai) of M the fibre of the symbol bundle is spanned by the matrices B λ = (∂F λ /∂pai )(q). Thus, the symbol of the Pfaffian system (63) associated to the P.D.E. system (62) agrees with the classical definition of the symbol of such a system. In general, we have chosen our notations for a linear Pfaffian system so that their structure equations look like the structure equations of the special system (63). In this regard, we offer without proof the following easy


117

Proposition 5.10. A linear Pfaffian system is locally equivalent to the Pfaffian system (63) arising from a P.D.E. system (62) if, and only if, the Frobenius condition dJ ≡ 0 mod {J} is satisfied. It is well known that a P.D.E. system may have compatibility conditions obtained from the equality of mixed partials, and we shall find the expression of these conditions for a general linear Pfaffian system. More specifically, we consider the compatibility conditions for the affine linear equations (65)

(Aaεi (x)pεj − Aaεj (x)pεi ) + caij (x) = 0

whose solutions give the integral elements Gx (I, Ω) lying over a point x ∈ M . The compatibility conditions for this system of linear equations in the pεi may lead to relations on the Aaεi (x) and caij (x). These are called integrability conditions. Their presence means that the set G(I, Ω) of integral elements of (I, Ω) projects onto a proper subset of M , and we should restrict our consideration to this subset. As the following simple example shows, the presence of integrability conditions is an important phenomenon for “over-determined” systems of partial differential equations, and usually imposes strong restrictions on the solution. Example 5.11. In the (x, y, z)-space consider the system of P.D.E.’s of the first order: zy = B(x, y, z). zx = A(x, y, z), This is equivalent to the differential system (66)

θ = dz − Adx − Bdy = 0,

also in (x, y, z) space and with the independence condition (67)

dx ∧ dy = 0.

In the above notation we have rank I = 1, rank J = 3 and thus J = T ∗ M , and at each point of M there is a unique 2-plane (66) satisfying the independence condition (67). The condition that this 2-plane be an integral element is that dθ = −dA ∧ dx − dB ∧ dy restrict to zero on it. Working this out gives (68)

Ay + Az B = Bx + Bz A,

which is the usual integrability condition. If it is not identically satisfied, there are two cases: a) The relation (68) does not involve z and is therefore a relation between x, y, so that the system has no integral manifold satisfying the independence condition (67); b) The relation (68) gives z as a function of x, y, which is then the only possible solution, and thus the equation has a solution or not depending on whether it is satisfied or not by this function.

118


Example 5.12. In an open set U ⊂ Rn (n ≥ 2) with coordinates x1 , . . . , xn , we assume given a smooth 2-form ϕ = 12 ϕij dxi ∧ dxj , ϕij + ϕji = 0, and consider the equation dη + ϕ = 0 for a 1-form η. When written out, this equation becomes the P.D.E. system ∂ηi /∂xj − ∂ηj /∂xi + ϕij = 0. The associated exterior differential system is defined on the submanifold M of (xi , ηi , pij ) space by the equations ⎧ ⎪ ⎨ pij − pji + ϕij = 0 θi = dηi − pij dxj = 0 ⎪ ⎩ 1 dx ∧ · · · ∧ dxn = 0.

(69)

We seek an integral element E ⊂ Tq M defined by dpij − pijk dxk = 0

(70) together with

dθi |E = 0 (dpij − dpji + dϕij )|E = 0.

The first of these equations gives (71)

pijk = pikj ,

and using (70) the second equations give (72)

pijk − pjik + ∂ϕij /∂xk = 0.

It is an elementary consequence of (71) and (72) that dϕ = 0. In other words, the necessary and sufficient condition that the Pfaffian system (69) have an integral element lying over each point of M is that dϕ = 0. This again illustrates our assertion that the compatibility conditions for the equations (65) are integrability conditions; more precisely, they are first order integrability conditions. We will now see how these integrability conditions are reflected in the structure equations of a linear Pfaffian differential system. Referring to the proof of Proposition 5.2 above, we assume that θ1 , . . . , θs0 , ω1 , . . . , ωn , π 1 , . . . , π t is a local coframe for M adapted to the filtration I ⊂ J ⊂ T ∗ M . This coframe is defined up to an invertible linear transformation (47). The linearity of the Pfaffian system is expressed by the equation (51), i.e., by the absence of π δ ∧ π ε terms in the dθa ’s. By abuse of notation we shall write the system as ⎧ a ⎪ ⎨θ =0 dθa ≡ Aaεi π ε ∧ ωi + 12 caij ωi ∧ ωj mod {I} (73) ⎪ ⎩ Ω = ω1 ∧ · · · ∧ ωn = 0.


119

Under a substitution (47) with only the diagonal blocks being non-zero—i.e., with Ciε = 0 (the θ-terms don’t matter because of the congruence in the 2nd equation above)—we see that caij transforms like a section of I ∗ ⊗ Λ2 (J/I). Under a substitution (47) with the diagonal blocks being the identity, i.e., given by (74)

π ε → π ε + pεi ωi ,

we have that (75)

caij → caij + (Aaεj pεi − Aaεi pεj ).

In intrinsic terms, we have a mapping (76)

π : J ⊥ ⊗ J/I → I ∗ ⊗ Λ2 (J/I)

induced by (60) and given in the above bases by pεi → (Aaεj pεi − Aaεi pεj ), and the caij give a section of I ∗ ⊗ Λ2 (J/I)/image π, where we now assume that the mapping (76) has locally constant rank. Definition 5.13. We denote by [c] the section of the bundle I ∗ ⊗ Λ2 (J/I)/image π, given in bases by {caij } modulo the equivalence relation (75). Then [c] is called the torsion of the linear Pfaffian system (73). From the proof of Proposition 5.7 we have the Proposition 5.14. The necessary and sufficient condition that there exist an integral element of (73) over a point x ∈ M is that the torsion [c](x) = 0. By our discussion above, we see that the torsion reflects the 1st order integrability conditions in the Pfaffian system (73). For this reason, we shall sometimes say that the integrability conditions are satisfied rather than the torsion vanishes. On the other hand, assuming always that the mapping (76) has constant rank, we see that the vanishing of the torsion in an open neighborhood U of a point x ∈ M is equivalent to being able to make a smooth substitution (74) in U such that the caij = 0 in (73). For this reason, we shall sometimes say that the torsion may be absorbed (by a substitution (74)) rather than the torsion vanishes. In summary, we have that the satisfaction of the integrability conditions for (73) is expressed by being able to absorb the torsion. There is an alternate way of writing (73), called the dual form, that is especially useful in computing examples. To explain it we set (77)

πia = Aaεi π ε + caji ωj

so that the second equation in (73) becomes (78)

dθa ≡ πia ∧ ωi

mod {I}.

The 1-forms πia are not linearly independent modulo J, but are subject to the relations (79)

Baλi πia ≡ Cjλ ωj

mod {I},

120


where by (77) Cjλ = Baλi caji .

(80)

Here, we are working over an open set U ⊂ M and omitting reference to the point x ∈ M , and the Baλi give a basis for symbol bundle over U . Summarizing, the dual form of the structure equations (73) is ⎧ a θ =0 ⎪ ⎪ ⎪ ⎨ dθa ≡ π a ∧ ωi mod {I} i (81) λi a ⎪ π ≡ Cjλ ωj mod {I} B ⎪ ⎪ a i ⎩ Ω = ω1 ∧ · · · ∧ ωn = 0. Proposition 5.15. Assuming that the operator π in (76) has constant rank, the following are equivalent: (i) the space G(I, Ω) of integral elements surjects onto M ; (ii) locally, we may choose the π ε so that caij = 0 in (73); (iii) locally, we may choose the πia so that Cjλ = 0 in (81). Proof. We have proved that (i) ⇒ (ii) ⇒ (iii) above (see (80) for (ii) ⇒ (iii)). Assuming (iii), we consider the family of n-planes a θ =0 πia = 0. By the third equation in (81), these n-planes are well-defined, and by the first and second equations there they are integral elements. If any of the equivalent conditions in the proposition are satisfied, we shall say that the integrability conditions are satisfied or that the torsion may be absorbed. Suppose that (i) in the proposition is satisfied. Then we may choose our coframe θ1 , . . . , θs0 , ω1 , . . . , ωn , π 1 , . . . , π t so that the structure equations (73), (81) became respectively ⎧ a ⎪ ⎨θ =0 (82)

(83)

⎪ ⎩

dθa ≡ Aaεi π ε ∧ ωi mod {I} Ω = ω1 ∧ · · · ∧ ωn = 0

⎧ a θ =0 ⎪ ⎪ ⎪ ⎨ dθa ≡ π a ∧ ωi mod {I} i λi a ⎪ π ≡ 0 mod {I} B ⎪ a i ⎪ ⎩ Ω = ω1 ∧ · · · ∧ ωn = 0.

These are the forms of the structure equations that we shall use in examples where the torsion is absorbed. Referring to structure equations (73), we have discussed the tableau and torsion of a linear Pfaffian system. We will now express Cartan’s test for involution in terms of these invariants. For this we work in a neighborhood U of a point x ∈ M (1) and assume that the quantities dim Ax , si (x), dim Ax , and rank π x in (76) all are constant.


121

Theorem 5.16. The linear Pfaffian system (I, Ω) is involutive at x ∈ M if, and only if, (i) the torsion vanishes in U (ii) the tableau Ax is involutive. Proof. We may replace U by M , and then by Proposition 5.14 the vanishing of the torsion is equivalent to the surjectivity of the mapping G(I, Ω) → M . By Proposition 5.7 we then have that dim G(I, Ω) = dim M + dim A(1) x . On the other hand, using the structure equations (83) we see first that the equations

θa (x) = 0 πia (x) = 0

define an integral element E ⊂ Tx M having a basis ei where ωj (x), ei = δij , and secondly the proof of (36) in section 3 above shows that the rank of the polar equations of Ek = span{e1 , . . . , ek } is given by s0 + s1 + · · · + sk . On the other hand, the Cartan characters rk and sk associated to E in terms of the dimensions of the polar spaces H(E k ) are given for k ≥ 0 by the relations (84)

dim H(E k ) = rk+1 + k + 1 sk = rk − rk+1 − 1 ≥ 0.

We set r0 = m = dim M , reflecting the assumption that there are integral elements over each point. Then r0 − r1 − 1 = rank Ix = s0 , so that our notations are consistent. From (84) we infer that, for 1 ≤ k ≤ n − 1, (85)

rk+1 + k + 1 = n + t − (s0 + s1 + · · · + sk ).

Subtracting these equations for k and k − 1 and using the second equation in (84) gives (86)

sk = sk

for k = 1, . . . , n − 1. This proves: For any integral element E, the characters sk = sk (E) are equal to the reduced characters sk . In particular, the sk are the same for all integral elements E lying over a fixed point x ∈ M . The inequality in Cartan’s test given by Theorem 1.11 in Chapter III is then (87)

dim A(1) x ≤ s1 + 2s2 + · · · + nsn ,

and, by the Definition 3.7 above, equality holds if, and only if, Ax is involutive.

122


To conclude this section, we want to give a practical method for computing the si so that one can, with relative ease, check for equality in (87). For a 1-form ϕ we set ϕ = ϕ(x) modulo Jx = ϕ(x) modulo {θa (x), ωi (x)}, and we shall omit reference to the point x ∈ M . We define the tableau matrix by ⎡ 1 ⎤ π 1 . . . π 1n ⎢ .. ⎥ . (88) π = ⎣ ... . ⎦ s0 π 1 . . . π sn0 Then, assuming that the ωi are chosen generically, we have again from the proof of Proposition 3.8 above that number of independent 1-forms π ai . (89) s1 + · · · + sk = in the first k columns of π In practice, this equality will allow us to determine the si by “eyeballing” the tableau matrix (88). For an illustration of the use of the tableau matrix, we shall put it in a normal form and use this to give an especially transparent proof of Cartan’s test. We assume that the torsion has been absorbed so that the structure equations (83) and relations (86), (89) hold. Then, amongst the 1-forms π a1 exactly s1 are linearly independent. (We remind the reader that we are omitting reference to s the point x.) We may then assume that θ1 , . . . , θs0 are chosen so that π 11 , . . . , π11 are independent. Then all of the forms πia for a > s1 are linear combinations s of π 11 , . . . , π11 , since otherwise we could choose ω1 , . . . , ωn so that at least s1 + 1 of the forms π a1 were linearly independent in contradiction to (89). Having said s s this, among the forms π 12 , . . . , π 21 exactly s2 are independent modulo π 11 , . . . , π 11 . s We may assume that θ1 , . . . , θs1 have been chosen so that π 12 , . . . , π22 are linearly s independent modulo π 11 , . . . , π 11 . We note that s2 ≤ s1 , since otherwise s1 would not be the number of independent π a1 for a generic choice of ω1 , . . . , ωn . Continuing in this way, we may choose θ1 , . . . , θs0 so that the tableau matrix looks like π 11 .. . .. . s π 11 ∗ ∗

(90)

π 12 . . . π 1n .. s . π nn s

π 22 ∗ ∗ ∗ ...

∗ ∗ ∗ ∗

with the property that: for b > sk the form π bk is a linear combination of the forms π ai where i ≤ k, a ≤ si .6 We also note that s1 ≥ s2 ≥ · · · ≥ sn . 6 Actually,

s

we have proved that the πbk for b > s1 are linear combinations of π 11 , . . . , π11 ; the s1

s2

πbk for s2 < b ≤ s1 are linear combinations of π 11 , . . . , π 1 , π 12 , . . . , π2 ; the πbk for s3 < b ≤ s2 are s

s

s

linear combinations of π 11 , . . . , π 11 , π 12 , . . . , π 22 , π 13 , . . . , π 33 ; and so forth.


123

Definition 5.17. The forms π ai where a ≤ si are called the principal components. Thus the principal components are independent and span the space of the π ai ’s. We will now derive Cartan’s test. Integral elements are defined by linear equations θa = 0 πia − paij ωj = 0 where paij = paji and Baλi paij = 0 by (83). Clearly it suffices to consider only those linear equations πia − paij ωj = 0,

(91)

a ≤ si

corresponding to the principal components, as the remaining linear equations are consequences of these by writing πkb

=

linear combination of the πia for i ≤ k and a ≤ si together with the ωj and θb .

Since paij = paji the integral element (46) is determined by the quantities paij , a ≤ min si , sj or equivalently by the quantities (92)

paij ,

i ≤ j,

a ≤ sj .

For j = 1 we have the pa11 for a ≤ s1 . For j = 2 we have the pa12 and pa22 for a ≤ s2 . For j = 3 we have the pa13 , pa23 , and pa33 for a ≤ s3 . Continuing in this way, we see that there are at most (93)

s1 + 2s2 + 3s3 + · · · + nsn

independent quantities (92), and this is the inequality in Cartan’s test. Suppose that equality holds, and consider integral elements of dimension p that satisfy the additional equations ωp+1 = · · · = ωn = 0. These integral elements are given by equations (91) where we set ωp+1 = · · · = ωn = 0, from which it follows that they are uniquely determined by the quantities (92) where j ≤ p. If the space of integral elements is of dimension equal to (93), the paij in (92) can be freely specified. From this it follows that every p-dimensional integral element given by ωp+1 = · · · = ωn = 0 extends to a (p + 1)-dimensional integral element given by ωp+2 = · · · = ωn = 0, and in fact does so in (p + 1)sp -dimensional ways. Thus, we have a C-regular flag, which means that the system is involutive. Remark. To understand the relation between this discussion and the proof of Cartan’s test in Chapter III, we assume as above that E n is defined by the equations θa = πib = 0 and E k ⊂ E n by the additional equations ωk+1 = · · · = ωn = 0. In the filtration (94)

E0 ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En ⊂ H(En ) ⊂ . . . ⊂ H(E2 ) ⊂ H(E1 ) ⊂ H(E0 ) ⊂ T

124


we have H(E0 ) = {θb = 0} H(E1 ) = {θb = 0, π1a = 0 for a ≤ s1 } H(E2 ) = {θb = 0, π1a = 0 for a ≤ s1 and π2b = 0 for b ≤ s2 } .. .

.. .

H(En ) = {θb = 0, πia = 0 for a ≤ si }. From this it is clear that the normal form (90) is simply the implication on the tableau of chosing the πia adapted to the filtration (94) in the manner just explained. A very useful insight into involutivity is to express its consequences on the symbol relations when the tableau matrix is in the normal form (90). The general case is called the Guillemin normal form, which will be further discussed at the end of Chapter VIII. Here we will first take up the special case when the system is involutive and s1 = s0 ,

(95)

s2 = · · · = sn = 0.

Thus the principal components are the π1a , and using the additional index range 2 ≤ ρ, σ ≤ n we will have relations a b πρa ≡ Cρb π1 ,

(96)

where the congruence is modulo the θa ’s and ωi ’s. These are a complete set of symbol relations, and setting a Cρ = Cρb we will prove that: (97) (98)

A tableau satisfying (95) with symbol relations (96) is involutive if, and only if, the commutation relations [Cρ, Cσ ] = 0 are satisfied.

Proof. Among the equations πia − paij ωj = 0 that define integral elements, by (95) those for i ≥ 2 are consequences of those for i = 1. Thus, by Cartan’s test the pa11 may be freely specified and then the (99)

pai1 = pa1j paρσ = paσρ


125

are determined. From (96) we have a b paρi = Cρb p1i .

Thus a b paρ1 = Cρb p11

(100) and

a b paρσ = Cρb p1σ .

(101)

Using the first equation in (99) and (100) in (101) gives a b c paρσ = Cρb Cσc p11 ,

and then the second equation in (99) gives a b a b (Cρb Cσc − Cσb Cρc )pc11 = 0.

Since the pc11 may be freely specified, we conclude (98). Reversing the argument gives our assertion. A slightly more general case arises when we assume that (102)

s1 = · · · = sl = s0 ,

sl+1 = · · · = sn = 0.

With the ranges of indices 1 ≤ λ ≤ l,

l+1 ≤ ρ ≤n

the complete symbol relations are λa b πρa ≡ Cρb πλ .

For any ξ = (ξ1 , . . . , ξl ) we define λa Cρ (ξ) = Cρb ξλ ,

and the the same proof gives: (103)

A tableau satisfying (102) is involutive if, and only if, for all ξ, the commutation relations

(104)

[Cρ (ξ), Cσ (ξ)] = 0, are satisfied.

Although we shall not completely write out the Guillemin normal form here, we will refine the normal form (90) and indicate how this in fact leads to a generalization of Guillemin’s result.

126


We first claim that: (∗)

In the involutive case, we may choose θ1 , . . . , θs0 so that π ai = 0 for a > s1 ,

1 ≤ i ≤ n.

Proof. We may assume that the torsion is absorbed, i.e. that (83) holds, and then the 1-forms θa , ωi , πia where a ≤ si are linearly independent. In the tableau matrix πia without reducing modulo J, even though this is not intrinsic, we have s1 = number of linearly independent 1-forms π1a . Moreover, if we then set all π1a = 0 and denote by ϕ˜ the restriction of any 1-form ϕ to this space, the tableau with matrix ˜ πρa , 2 ≤ ρ ≤ n and 1 ≤ a ≤ s0 is again involutive. We choose θ1 , . . . , θs0 so that π1a = 0,

a > s1 ,

and will show that, as a consequence of involutivity, all the remaining πρa = 0,

a > s1

and

2 ≤ ρ ≤ n.

For this we recall from the argument for Cartan’s test given below Definition 5.12 that integral elements are defined by equations (i)

πia − paij ωj = 0, paij = paji ,

and that the pa11 , a ≤ s1 , may be freely specified. We now use the additional range of indices 1 ≤ λ ≤ s1 . Then we have equations a πρa = Bρλ π1λ ,

2 ≤ ρ ≤ n.

Using (i) these give equations a λ j paρj ωj = Bρλ p1j ω a λ ⇒paρ1 = Bρλ p11 .

But π1a = 0 for a > s1 gives a λ 0 = pa1ρ = paρ1 = Bρλ p11 , a = 0. and the only way the pλ11 can be freely specified is if all Bρλ

§6. Prolongation

127

We now may repeat (i) for the involutive tableau ˜ πρa to conclude that, with a 1 s1 a suitable choice of θ , . . . , θ , all π ˜ρ = 0 for s2 < a ≤ s1 . This is equivalent to s2 < a ≤ s1 .

πρa ≡ 0 mod {π1b },

Continuing in this way we may assume that the tableau matrix has the following form

(ii)

π11 .. . .. . .. .

π21 .. . .. .

π31 .. . s

π3 3

s

π2 2

s π1 1

...

Ψ2 Ψ1

Ψ0 where

⎧ Ψ0 = 0 ⎪ ⎪ ⎪ ⎪ 1 1 ⎪ ⎨ Ψ1 ≡ 0 mod {π1 , . . . , πs1 } Ψ2 ≡ 0 mod {π11 , . . . , πs1 , π12 , . . . , πs2 } ⎪ ⎪ 1 2 ⎪ ⎪ ⎪ ⎩ .. .

The Guillemin normal form arises by writing

(iii)

⎧ Ψ = C11 π1 ⎪ ⎨ 1 Ψ2 = C21 π1 + C22 π2 ⎪ ⎩ .. .

where π1 , π2, . . . are the columns in the tableau matrix. We may then repeat the argument that gave (98) to deduce a set of quadratic conditions on the symbol relations (iii) that are necessary and sufficient in order that the tableau be involutive.7 We shall not pursue this further here.

§6. Prolongation. Let I be an exterior differential system on a manifold M . The first prolongation will be a linear Pfaffian system (I (1), Ω) with independence condition on a manifold M (1). Roughly speaking, the first prolongation is obtained by imposing the first 7 Actually, (ii) is a slight refinement of the Guillemin normal form, which more closely corresponds to the normal form (90). The way to understand (ii) is as follows: Given a generic flag V1 ⊂ V2 ⊂ · · · ⊂ Vn−1 ⊂ V , dim Vi = i, this determines a flag W ⊃ W1 ⊃ W2 ⊃ · · · ⊃ Wn−1 , dim Wi = si together with a set of matrices Cij , i ≥ j, where Cij is an (si − si+1 ) × (sj − sj+1 ) matrix and where a set of quadratic relations is imposed on the Cij . Counting the number of independent equations would allow one to compute the dimension of the space of involutive tableaux with fixed characters, which to our knowledge has never been done.

128


order integrability conditions on the original system. The prolongation (I (1) , Ω) of an exterior differential system (I, Ω) with independence condition will also be defined, and then the higher prolongations are defined inductively by I (0) = I and (I (q+1) , Ω) = 1st -prolongation of (I (q) , Ω). The three basic properties of prolongation may be informally and imprecisely stated as follows: (105)

The integral manifolds of (I, Ω) and (I (1), Ω) are locally in one-to-one correspondence.

(106)

If (I, Ω) is involutive, then so is (I (1), Ω).

(107)

There exists a q0 such that, for q ≥ q0 , (I (q) , Ω) is involutive.

This last property includes the possibility that the manifolds M (q) are empty— this is the case when there are no integral manifolds of (I, Ω). As a consequence of (105) and (107) we may (again imprecisely) say that every integral manifold of a differential system is an integral manifold of an involutive exterior differential system. We will now give the definition of (I (1) , Ω) in a special case; the general definition will be taken up in Chapter VI. For this we assume that the variety Gn (I) of ndimensional integral elements is a smooth submanifold of Gn (T M ) whose defining equations are derived from I as explained in Chapter III (cf. Proposition 1.4 and Definition 1.7 there). Thus (108)

Gn (I) = {E ∈ Gn (T M ) : ϕ|E = 0 for all ϕ ∈ I}

should be a regularly defined submanifold of Gn (T M ). Definition 6.1. The first prolongation (I (1) , Ω) is defined to be the restriction to Gn (I) of the canonical system (L, Φ) on Gn (T M ). To see what prolongation looks like in coordinates, we suppose that (x1 , . . . , xn, y1 , . . . , ys ) is a coordinate system on M and we consider the open set U ⊂ Gn (T M ) given by tangent planes E such that dx1 ∧ · · · ∧ dxn|E = 0. These tangent planes are defined by equations (see the discussion in Example 5.4 above) (109)

θσ = dyσ − pσi dxi = 0,

and then (xi , yσ , pσi ) forms a local coordinate system on Gn (T M ). In this open set, the canonical system L is generated by the 1-forms θσ together with their exterior derivatives and the independence condition is given by Φ = dx1 ∧ · · ·∧ dxn . When written out in this coordinate system, the equations (108) become a system of equations (110)

Fϕ(x, y, p) = 0,

ϕ ∈ I,

§6. Prolongation

129

where ϕ|E = Fϕ(x, y, p)Φ for E given by (109). Our assumption is that these equations regularly define a submanifold in (x, y, p) space, and then (I (1), Ω) is the restriction to this submanifold of the canonical system. It is clear that (111)

The first prolongation is a linear Pfaffian differential system with independence condition.

Moreover, property (105) above is also clear from the fact (cf. example 1.5) that the integral manifolds of (L, Φ) are locally the canonical liftings of smooth mappings f : N → M . Properties (106) and (107) are more subtle and will be taken up later. If we begin with a differential system (I, Ω) with independence condition, then assuming as above that G(I, Ω) is a smooth submanifold, (I (1) , Ω) is defined to be the restriction to G(I, Ω) of the canonical system on Gn (T M ). Example 6.2. We will work out the structure equations for the first prolongation of a linear Pfaffian system for which the torsion vanishes. First we make a general observation. Using the above notation, suppose that ϕ = fσ (x, y)dyσ − gi (x, y)dxi is a 1-form in I where some fσ = 0. The condition that ϕ vanish on the integral element (109) is fσ (x, y)pσi = gi (x, y).

(112)

We let π : M (1) → M be the projection, and on M (1) we consider the 1-form π ∗ ϕ = fσ (x, y)dyσ − gi (x, y)dxi = fσ (x, y)(dyσ − pσi dxi ) by (112) = fσ θσ by (109). In summary: (113)

Let π : M (1) → M be the canonical projection, and let I1 denote the differential ideal generated by the 1-forms in I. Then π ∗ I1 ⊆ I (1). In particular, if I is a Pfaffian system then π ∗ I ⊂ I (1) .

From now on we will usually omit the π ∗ ’s.

130


Returning to our example, we suppose I to locally have the structure equations (83) above. Then M (1) → M is a bundle of affine linear spaces whose associated (1) vector bundle has fibre Ax over x ∈ M . Integral elements of (I, Ω) are defined by (114)

θa = 0 θia = πia − paij ωj = 0

where the equations (115)

paij = paji Baλi (x)paij = 0

are satisfied. From (113) it follows that I (1) is generated as a differential ideal by the 1-forms θa , θia (recall that we are omitting the π ∗ ’s). For the structure equations of I (1) we let I (1) ⊂ T ∗ M (1) be the sub-bundle whose sections are the 1-forms in I (1), so that locally I (1) = span{θa , θia }. We shall prove that: (116)

dθa ≡ 0

mod {I (1) }

(117)

a ∧ ωj dθia ≡ πij

mod {I (1)}

where (118)

a a πij ≡ πji mod {I (1) } a λ k Baλi πij ≡ Cjk ω mod {I (1) }.

Proof of (116). From (83) and (114) above dθa ≡ πia ∧ ωi mod {I} ≡ θia ∧ ωi mod {I}

(119)

≡ 0 mod {I (1) }. Proof of (117). We shall use the following variant of the Cartan lemma to be proved in Chapter VIII (cf. Proposition 2.1 and its corollaries in Chapter VIII): Let ω ˜ i be 2 linearly independent vectors in a vector space U and let ηi ∈ Λ U satisfy the exterior equation ˜ i = 0. ηi ∧ ω

(120) Then it follows that (121)

ηi = ηij ∧ ω ˜ j where ηij = ηji ∈ U.

§6. Prolongation

131

To apply this, we take the exterior derivative of (119) and use (116) and (114) to obtain 0 ≡ dπia ∧ ωi − πja ∧ dωj mod {I (1) } ≡ (dπia − paij dωj ) ∧ ωi mod {I (1) }. We take U to be a typical fibre of T ∗ M (1)/I (1) and ω ˜ i = ωi mod {I (1)}. Then from (120) and (121) we conclude that (122)

a ∧ ωj dπia − paij dωj ≡ ηij

mod {I (1) }

where a a ≡ ηij ηij

mod {I (1)}.

Now we have from the definition (114) and (122), dθia = dπia − paij dωj − dpaij ∧ ωj a ≡ πij ∧ ωj mod {I (1) }

where (123)

a a πij = −dpaij + ηij a a πij ≡ πji mod {I (1) }.

This gives (117) where the 1st equation in (118) is satisfied. To verify the second equation in (118) we let J (1) ⊂ T ∗ M (1) be the sub-bundle generated by I (1) and the values of the ωi (x). Then exterior differentiation of the second equation in (115) gives Baλi (x)dpaij ≡ 0 mod {J (1)} since dBaλi (x) ∈ Tx∗ M and the θa (x), πia (x), ωi (x) span Tx∗ M . From this and the 1st equation in (123) we have (dropping reference to x) a Baλi πij ≡0

mod {J (1)}

a since ηij ≡ 0 mod {J (1)}. This implies the 2nd -equation in (118).

Using (116)–(118) the structure equations of (I (1) , Ω) may be summarized as follows: (i) θa = 0 (ii) θia = 0 (124)

(iii) dθa ≡ 0 mod {I (1) } a ∧ ωj mod {I (1) } (iv) dθia ≡ πij a a = πji (v) πij a λ (vi) Baλi πij ≡ Cjk ωk mod {I (1) }

132


with the independence condition Ω = ω1 ∧ · · · ∧ ωn = 0. From this we conclude that (125)

(I (1) , Ω) is a linear Pfaffian system whose tableau over a point y ∈ M (1) is the 1st prolongation A(1) x of the tableau of (I, Ω) over x = π(y).

In other words, the tableau of the prolongation is the prolongation of the tableau. It follows also that the integral elements of (I (1), Ω) over a point y ∈ M (1) form (2) an affine linear space whose associated vector space is the 2nd prolongation Ax where x = π(y). In fact, referring to (124) there integral elements are given by a πij − paijk ωk = 0

where

paijk = pajik = paikj

(by (iv) and (v))

Baλi paijk

(by (vi)).

=

λ Cjk (y)

The homogeneous linear equations associated to the second of these equations define (2) Ax ⊂ W ⊗ S 3 V ∗ . We will return to these matters in Chapters V and VI. Finally, for use in Chapter VIII we want to prove the relation (126)

λ j Cjk ω ∧ ωk ≡ 0 mod {I (1)}.

Proof. From the equations ⎧ λi a ⎪ ⎨ Ba πi ≡ 0 mod I1 Baλi paij = 0 ⎪ ⎩ a θi = πia − paij ωj together with (113) above, we infer that Baλi θia ≡ 0 mod I1 Baλi dθia ≡ 0 mod I (1). Plugging this into equation (iv) in (124) above and using (vi) there gives (126). Example 6.3. We consider a 1 (127)

st

order P.D.E. system

F λ (xi , z a , ∂z a /∂xi ) = constant

(we do not specify what the constant is). On M = J 1 (Rn , Rs0 ) with coordinates (xi , z a, pai ) this P.D.E. system corresponds to the exterior differential system (I, Ω) generated by the 1-forms (i) dF λ (x, z, p) (128) (ii) θa = dz a − pai dxi

§7. Examples

133

and with independence condition Ω = dx1 ∧ · · · ∧ dxn . We want to see what, if any, P.D.E. system corresponds to the first prolongation (I (1), Ω). In fact, the prolongation of the P.D.E. system (127) is usually defined by introducing new variables pai for the derivatives ∂z a /∂xi and differentiating (127). Explicitly, it is the 1st order P.D.E. system for unknown functions z a , pai ⎧ a a i ⎨ pi = ∂z /∂x 1 (127 ) ∂F λ ∂F λ a ∂F λ ∂ ⎩ + p + (pa ) = 0, i = 1, . . . , n i ∂x ∂z a i ∂paj ∂xi j where F λ = F λ (x, z, p). Clearly the solutions of (127) and (1271 ) are in one-to-one correspondence (this is the reason for the constant in (127)), and we shall check that: (129)

(I (1), Ω) is the exterior differential system corresponding to the P.D.E. system (1271 ).

Proof. By definition, the exterior differential system corresponding to (1271 ) occurs ˜ ˜ of a jet space with coordinates (xi , z a, pa , q a, pa ), where M on a submanifold M i i ij has defining equations ⎧ a a ⎨ pi = q i (130) ∂F λ ∂F λ a ∂F λ a ⎩ + q + p = 0, ∂xi ∂z a i ∂paj ji where F λ = F λ (x, z, p). The differential ideal I˜ is generated by the restrictions to ˜ of the 1-forms M a θ = dz a − qia dxi θia = dpai − paij dxj . ˜ as being defined in Imposing the first equations in (130), we may think of M (xi , z a, pai , paij = paji ) space by the equations (131)

∂F λ ∂F λ a ∂F λ a + p + p =0 i ∂x ∂z a i ∂paj ij

˜ of the 1-forms and the ideal I˜ is generated by the restrictions to M a θ = dz a − pai dxi (132) θia = dpai − paij dxj . On the other hand, integral elements to (128) are defined by dpai − paij dxj = 0 subject to the conditions (from (ii) in (128)) −dpai ∧ dxi = −paij dxj ∧ dxi = 0, which implies that paij = paji , and (from (i) in (128)) ∂F λ i ∂F λ a i ∂F λ a i dx + p dx + p dx = 0. ∂xi ∂z a i ∂paj ij ˜ with M (1) and that, when Comparing with (131) we see that we may identify M (1) ˜ this is done, I = I .

134


§7. Examples. We will give some examples of differential systems in involution. Example 7.1 (Cauchy–Riemann equations). Let w = u + iv be a holomorphic function in the n complex variables z i = xi +

√ −1 yi ,

1 ≤ i ≤ n.

The Cauchy–Riemann equations can be written as the differential system θ1 = du − (pi dxi + qi dyi ) = 0 θ2 = dv − (−qi dxi + pi dyi ) = 0 in the space (xi , yi , u, v, pi, qi ) of 4n + 2 dimensions, the independence condition being Λi dxi ∧ dyi = 0. We have −dθ1 = dpi ∧ dxi + dqi ∧ dyi , −dθ2 = −dqi ∧ dxi + dpi ∧ dyi . To prove that the system is involutive we can proceed in one of the following two ways: 1) We search for a regular integral flag E 1 ⊂ E 2 ⊂ · · · ⊂ E n ⊂ E n+1 ⊂ · · · ⊂ E 2n such that E 2n is defined by dpi = (hij dxj + kij dyj ) dqi = (lij dxj + mij dyj ),

1 ≤ i, j ≤ n,

and E j , E n+j respectively by the further equations dxj+1 = · · · = dxn = 0, dy1 = · · · = dyn = 0, dyj+1 = · · · = dyn = 0. The conditions for E j to be integral are hij = hji , lij = lji , i ≤ j. The conditions for E n+j to be integral are −kij + lji = 0, mij + hji = 0, kij = kji, mij = mji .

§7. Examples

135

In both cases the equations are compatible, taking account of earlier equations expressing respectively, the conditions that E 1 ⊂ · · · ⊂ E j−1 and E 1 ⊂ · · · ⊂ E j−1+n are integral flags. Hence the system is in involution. 2) We apply Cartan’s test. The tableau matrix given by (88) in the previous section is (we omit the bars over the πia ) Π=

dp1 . . . dpn dq1 . . . dqn −dq1 −dqn dp1 . . . dpn

.

It is easily checked that s1 = · · · = sn = 2, sn+1 = · · · = s2n−1 = r2n = 0.

Both sides of the inequality (87) are equal to n(n + 1), and the system is in involution. Example 7.2 (Partial differential equations of the second order). As discussed above, the basis of second-order P.D.E.’s is the differential system θ = dz − pi dxi = 0 θi = dpi − pij dxj = 0; pij = pji , 1 ≤ i, j ≤ n in the space (xi , z, pi, pij ) of dimension 1 2n + 1 + n(n + 1). 2 We have −dθ = dpi ∧ dxi ≡ 0,

mod θi ,

−dθi = dpij ∧ dxj . To illustrate the scope of our concept of involutiveness, we wish to remark that the system is involutive, if there is no relation between the variables. In fact, define the admissible integral elements by dpij = pijk dxk ,

1 ≤ i, j, k, l ≤ n.

Then we have pijk = pikj , and therefore pijk is symmetric in any two of its indices. Thus dim Gx (I, Ω) = n(n + 1)(n + 2)/6. On the other hand, consider the tableau matrix (where we again omit the bars) ⎛

⎞ dp11 . . . dp1n ⎠. π = ⎝ ... dpn1 . . . dpnn We find

s1 = n, s2 = n − 1, . . . , sn−1 = 2, rn = sn = 1.

136


The involutiveness follows from the identity

i(n − i + 1) =

1≤i≤n

1 n(n + 1)(n + 2). 6

Suppose now there is one equation F (xi , z, pi, pij ) = 0. Then its differential gives ∂F dpij + · · · = 0, ∂pij and the dpij are linearly dependent. An advantage in using differential forms is that we can choose a basis to write this equation in a simple form. We put ωi = dxi , and apply the substitution ˜ i, ω ˜ i = uij ωj , ωj = vij ω θ˜i = vij θj , θj = uij θ˜i , where (uji ), (vij ) are inverse matrices to each other so that θi ∧ ωi = θ˜i ∧ ω ˜ i. Then we have dθ˜i ≡ π ˜il ∧ ω ˜l where π ˜ il = π ˜li = − or −dpjk =

mod {θ, θj }

dpjk vij vlk ,

uij ulk π˜il .

Suppose we make the non-degeneracy assumption det(∂F/∂pjk ) = 0. Then we can choose uij so that ∂F εi = ±1, i = l, i l uj uk = ∂pjk 0, i = l. This gives we put

εi π ˜ii +

≈

Ck ω ˜ k ≡ 0 mod {θ, θj } ≈

π ii = π ˜ii + εi Ci ω ˜ k , π ij = π˜ij , i = j, ≈

and we absorb θ, θj into π ij . By dropping the tildes, we arrive at the normal form (133)

dθi ≡ πij ∧ ωj mod {θ, θj }, εi πii = 0.

§7. Examples

137

From this the involutiveness of the system follows immediately. In fact, we put ⎛ ⎞ π11 . . . π1n ⎠ π = ⎝ ... πn1 . . . πnn where πij = πji ,

εi πii = 0.

We find s1 = n, s2 = n − 1, . . . , sn−1 = 2, sn = 0 so that s1 + 2s2 + · · · + (n − 1)sn−1 + nsn =

1 n(n + 1)(n + 2) − n. 6

On the other hand, the admissible n-dimensional integral elements on which θ = θj = 0 are given by πij = lijk ωk

(134) where

lijk = ljik = likj ,

εi liik = 0.

Its space has the dimension 16 n(n + 1)(n + 2) − n. Hence the system is in involution. However, such a result, that the system arising from a non-degenerate secondorder P.D.E. is in involution, does not seem to be exciting. To get an idea of the meaning of involutiveness, we will study a system of q equations (135)

F λ (xi , z, pi , pij ) = 0,

1 ≤ λ, µ ≤ q.

For q > 1, the system is “over-determined”, and we should expect strong conditions for it to be involutive. First the integrability conditions have to be satisfied. These can be expressed in terms of the functions F λ . We suppose this to be the case and proceed to study the conditions for involutivity in terms of the tableau. By the structure equation (83) in the previous section there exist πij such that dθi ≡ πij ∧ ωj mod {θj }, (136)

πij = πji , λ λ λ Bij πij = 0 mod {θj }, Bij = Bji .

In the matrix π we have s1 ≤ n, s2 ≤ n − 1, . . . , sn−2 ≤ 3, s1 + · · · + sn−1 ≤ 12 n(n + 1) − q, s1 + · · · + sn−1 + sn = 12 n(n + 1) − q.

138


For the system to be in involution we must have dim Gx (I, Ω) = s1 + 2s2 + · · · + (n − 1)sn−1 + nsn 1 =n n(n + 1) − q − (n − 2)s1 − · · · − sn−2 − s1 − · · · − sn−1 2 1 ≥ (n − 1) n(n + 1) − q − (n − 2)n − · · · − 1 · 3 2 1 (n − 1)(n2 + 4n + 6) − (n − 1)q. 6 On the other hand, for (134) to be an integral element we have lijk = ljik = likj , λ (137) Bij lijk = 0. =

The lijk , being symmetric, has 16 n(n+1)(n+2) components. Hence, in the involutive case, the number of linearly independent equations in the system (137) is 1 n(n + 1)(n + 2) − dim Gx(I, Ω) ≤ (n − 1)q + 1. 6 This condition can be reformulated as follows: We introduce the quadratic forms λ i j B λ = Bij ξξ .

The nq cubic forms ξk Bλ satisfy a linear relation if and only if (138)

mkλ ξ k B λ = 0,

λ λ λ (Bij mkλ + Bjk miλ + Bki mjλ ) = 0.

λ

The quantities defined by λ λ λ λ Bijk,l = Bij δkl + Bjk δil + Bki δjl

are the elements of an 16 n(n+1)(n+2)×nq matrix and are the coefficients of each of the systems (137) and (138). Hence the two systems have the same rank. It follows that the cubic forms ξ k B λ satisfy at least q − 1 independent linear relations.8 Consider the case q = 2. The above conclusion implies that l1 (ξ)B 1 + l2 (ξ)B 2 = 0, where l1 (ξ) and l2 (ξ) are linear forms. Hence B 1 and B 2 have a common linear factor. By a change of coordinates we have two cases: a) B 1 = ξ 1 ξ 2 , B 2 = ξ 1 ξ 3 ; b) B 1 = (ξ 1 )2 , B 2 = ξ 1 ξ 2 . For the corresponding differential systems we have the normal forms: a) π12 = π13 = 0; b) π11 = π12 = 0. We state our results as a proposition following from Theorem 5.16: 8 In §7 of Chapter VIII this will be generalized to P.D.E. systems of any order—cf. (146)–(152) there.

§7. Examples

139

Proposition 7.3. For a system of two second-order P.D.E.’s to be in involution it is necessary and sufficient that: 1) The integrability conditions be satisfied; 2) The symbols, as quadratic forms B 1 and B 2 , have a common linear factor. Proof. The sufficiency follows from verifying the involutiveness conditions when the system is in the normal forms a) or b). We leave this to the reader. Example 7.4. Consider the following equations which play a role in the old theory of matter and gravitation (see Cartan and Einstein [1979], p. 33): ∂Xi ∂Xj ∂Xi − = 0, = −4πρ j i ∂x ∂x ∂xi ∂ρ ∂(ρui ) =0 + ∂t ∂xi ∂ui ∂ui 1 ≤ i, j ≤ 3. + uj j = Xi , ∂t ∂x Here xi are the space coordinates, t is time, and ui and Xi are components respectively of the velocity and acceleration vectors, while ρ is the density of matter. This is a system of 8 equations in 4 independent variables xi , t, and 7 dependent variables ui , Xi , ρ. It is thus an overdetermined system. This system is involutive. To make the ideas more clear we will consider the simpler system (139)

∂Xi ∂Xj ∂Xi − = 0, = −4πρ, 1 ≤ i, j, k ≤ 3, j i ∂x ∂x ∂xi i

where Xi , ρ are functions of x1 , x2 , x3 , and ρ is given. We shall show that the system (139) is involutive. For this purpose we write it as a Pfaffian system Xij dxj dXi = (140) Xij = Xji , Xii = −4πρ, with the six Xij as new variables. The exterior derivatives of these equations give dXij ∧ dxj = 0 dXii = −4πdρ. Consider admissible integral elements E 3 defined by Xijk dxk , Xijk = Xjik . dXij = To show the involutivity of the system it suffices to find in E 3 a regular integral flag E 1 ⊂ E 2 ⊂ E 3 , such that E 1 and E 2 are defined respectively by dx2 = dx3 = 0 and dx3 = 0. The condition for E 1 to be integral is (141a)

Xii1 = −4π

∂ρ . ∂x1

140


The condition for E 2 to be integral are, in addition to (141a), (141b)

Xi12 = Xi21 ,

Xii2 = −4π

∂ρ , 1 ≤ i ≤ 3. ∂x2

These equations can be solved in terms of Xij2 , so that E 1 is regular. To see whether E 2 is regular we consider the conditions for E 3 to be integral, which are (141c)

Xi13 = Xi31 , Xi23 = Xi32 ,

Xii3 = −4π

∂ρ . ∂x3

The first two equations imply X231 = X132 . But this is the first equation of (141b), with i = 3. Hence it is satisfied, and we see that (141c) are compatible as linear equations in Xij3 . Thus E 2 is regular, and so is the integral flag E 1 ⊂ E 2 ⊂ E 3 . This proves the involutivity of the system (139) or (140). The proof of the involutivity of the original system is exactly the same. It is only suggested that one take as the starting one-dimensional integral element the one defined by dt = 0. dx1 = dx2 = dx3 = 0, In [1953] Cartan proved that Einstein’s field equations for a unified field theory based on distant parallelism are involutive.9 They are a highly overdetermined system.

§8. Families of Isometric Surfaces in Euclidean Space. In Chapter III we gave a proof of the Cartan–Janet isometric imbedding theorem. For two dimensions it says that an analytic Riemannian manifold of two dimensions can be locally isometrically imbedded in the 3-dimensional space E 3 . By the discussion in that chapter, the imbedding is not unique. The data needed to specify it uniquely will be discussed in the next chapter. In any case, we can say that it is not rigid, meaning that there is a surface isometric without being congruent to it. We therefore try to impose further conditions. The two natural conditions are: A) preservation of the lines of curvature; B) preservation of the principal curvatures. In each case we are led to an over-determined system, where the number of equations exceeds the number of unknown functions. Prolongation leads to new conditions and, in these two cases of geometrical significance, to very remarkable conditions. We shall state the two main results in this section. They are local results dealing with non-trivial families of isometric surfaces containing no umbilics, where a nontrivial family means a family which is not obtained from a given surface by a family of rigid motions. 9 The

specific references are Part III, Vol. 2, p. 1167–1185 and Part II, Vol. 2, p. 1199–1229.

§8. Families of Isometric Surfaces in Euclidean Space

141

Theorem 8.1. A non-trivial family of isometric surfaces of non-zero Gaussian curvature, preserving the lines of curvature, is a family of cylindrical molding surfaces. The cylindrical molding surfaces can be kinematically described as follows: Take a cylinder Z and a curve C on one of its tangent planes. A cylindrical molding surface is the locus described by C as the tangent plane rolls about Z. Theorem 8.2. A non-trivial family of isometric surfaces preserving the principal curvatures is one of the following: α) (the general case) a family of surfaces of constant mean curvature; β) (The exceptional case) a family of surfaces of non-constant mean curvature. They depend on six arbitrary constants and have the properties: β1 ) they are W -surfaces; β2 ) the metric (142)

dˆ s2 = (grad H)2 dx2 /(H 2 − K),

where ds2 is the metric of the surface and H and K are its mean curvature and Gaussian curvature respectively, has Gaussian curvature equal to −1. Our discussions in Chapter III contain the essence of a surface theory in E 3 , and we will summarize it in a form convenient for the present discussion as follows: We begin with the diagram P0 F ↓ π, M −→E 3

(143)

f

where P0 is the space of all orthonormal frames xe1 e2 e3 in E 3 , π(xe1 e2 e3 ) = x ∈ E 3 is the projection, f is the imbedding where we identify both the original and image point as x, F (x) = xe1 e2 e3 is a “lifting” of f satisfying the condition that e3 is the oriented unit normal at x, M being supposed to be oriented. The lifting F defines a family of frames over M which satisfy the equations dx = ω1 e1 + ω2 e2 , de1 = ω12 e2 + ω13 e3 (144)

de2 = −ω12 e1 + ω23 e3 de3 = −ω13 e1 − ω23 e2 .

Denoting by ( , ) the inner product in E 3 , the first and second fundamental forms of the surface M are I = ds2 = (dx, dx) = ω12 + ω22 , (145)

II = −(dx, de3 ) = ω1 ω13 + ω2 ω23 = l11 ω12 + 2l12 ω1 ω2 + l22 ω22 .

These depend only on the imbedding f, and are independent of the lifting F . The mean curvature and Gaussian curvature of M are respectively (146)

H=

1 2 . (l11 + l22 ), K = l11 l22 − l12 2

142


The eigenvalues of II with respect to I are the principal curvatures, and the eigen-directions, which are perpendicular, are the principal directions. If the lifting F is such that e1 , e2 are in the principal directions, II is diagonalized, i.e. l12 = 0. The point x is umbilic if the principal curvatures are equal; i.e., if H 2 − K = 0. We restrict ourselves to a neighborhood of M without umbilics, and choose e1 , e2 to be tangent vectors along the principal directions. Then the ω’s in (144) are all linear combinations of ω1 , ω2 and we set ω13 = aω1 , ω23 = cω2 (147)

ω12 = hω1 + kω2

Here a and c are the two principal curvatures; the mean curvature and the Gaussian curvature are now given by (148)

H=

1 (a + c), K = ac, 2

and the absence of umbilics is expressed by the condition a = c. Exterior differentiation of (144) gives the structure equations

(149)

dω1 = ω12 ∧ ω2 , dω2 = ω1 ∧ ω12 dωij = ωik ∧ ωkj , 1 ≤ i, j, k ≤ 3. k

The last equation, when written explicitly gives (150)

dω13 = ω12 ∧ ω23 , dω23 = ω13 ∧ ω12

(151)

dω12 = −Kω1 ∧ ω2 .

Equations (150) are called the Codazzi equations and equation (151) the Gaussian equation. We will use ω1 , ω2 to express the differential of any function on M , thus (152)

df = f1 ω1 + f2 ω2 ,

so that f1 , f2 are the “directional derivatives” of f. Using this notation and substituting the expressions in the first equations of (147) into (150), we get (153)

a2 = (a − c)h, c1 = (a − c)k.

We will use this formalism to study our isometry problems. To study problem A, let M ∗ be a surface isometric to M such that the isometry preserves the lines of curvature. Using asterisks to denote the quantities pertaining to M ∗ , we have ∗ ω1∗ = ω1 , ω2∗ = ω2 , ω3 = ω3∗ = 0, ω12 = ω12 ,

(154)

c ∗ ∗ = taω1 , ω23 = ω2 . ω13 t


143

The last two equations follow from the fact that M ∗ has the same Gaussian curvature as M at corresponding points. Equation (153) gives, when applied to M ∗ , c (ta)2 = (ta − )h, t c 1 ( c)1 = (ta − )k. t t

(155)

Comparison of (153) and (155) gives t1 = t(1 − t2 )ac−1 k,

(156)

t2 = −t−1 (1 − t2 )a−1 ch

or (156a)

tdt = t2 ac−1 kω1 − a−1 chω2 . 1 − t2

In fact, from now on we suppose t2 = 1, discarding the trivial case that M ∗ is congruent or symmetric to M . We set m = a−1 h,

(157)

n = c−1 k,

so that (158)

ω12 = mω13 + nω23

and define (159)

π1 = nω13 ,

π2 = mω23 .

Then (156a) can be written tdt = t2 π 1 − π 2 . 1 − t2

(156b) Its exterior differentiation gives (160)

t2 (dπ1 − 2π1 ∧ π2 ) = dπ2 − 2π1 ∧ π2 .

This equation, if not satisfied identically, completely determines t2 . On substituting into (156), we get conditions on the surfaces M , to which there exist isometric but not congruent or symmetric surfaces preserving the lines of curvature. The later are uniquely determined up to position in space. The most interesting case is when the equation (160) is identically satisfied, i.e., (161)

dπ1 = dπ2 = 2π1 ∧ π2 .

This leads to a non-trivial family of isometric surfaces preserving the lines of curvature.

144


In fact, substitution of (159) into (161) gives (dm − mnω13 ) ∧ ω23 = 0

(162)

(dn + mnω23 ) ∧ ω13 = 0.

We shall show that these equations imply mn = 0 or equivalently hk = 0. Equations (162) allow us to set dm = mnω13 + qω23 , (163)

dn = pω13 − mnω23 .

by (150) and (158) we have (164)

dω13 = ω12 ∧ ω23 = mω13 ∧ ω23 , dω23 = −ω12 ∧ ω13 = nω13 ∧ ω23 .

Taking the exterior derivative of (158) and using (163), (164), we get (165)

p − q + 1 + m2 + n2 = 0.

If m and n are considered as unknown functions, equations (163) and (165) give three relations between their derivatives. This primitive counting shows that the differential system is over-determined. To study our problem there is no other way but to examine the integrability conditions through differentiation of (163), (165). In this case the integrability conditions give a very simple conclusion. Exterior differentiation of (163) gives (166)

(dq + 2m2 nω13 ) ∧ ω23 = 0, (dp + 2mn2 ω23 ) ∧ ω13 = 0,

which allow us to set (167)

dp = rω13 − 2mn2 ω23 , dq = −2m2 nω13 + sω23 .

Differentiation of (165) then gives (168)

r = 2n(−2m2 − p), s = 2m(−2n2 + q).

As a result the prolongation “stabilizes” with (169)

dp = 2n(−2m2 − p)ω13 − 2mn2 ω23 , dq = −2m2 nω13 + 2m(−2n2 + q)ω23 .

Exterior differentiation of this equation and use of (163), (164), (169) gives mn(p − q + m2 + n2 ) = 0.


145

Comparing with (165), we get mn = 0, or equivalently hk = 0, as claimed above. We wish to describe these surfaces geometrically. Suppose k = 0. Then, by (163), (165), p = 0, q = 1 + m2 . It follows that the surfaces in question satisfy the equations

(170)

ω3 = 0, ω13 = aω1 , ω23 = cω2 , ω12 = hω1 , h h2 d = c 1 + 2 ω2 , a a ω1 ∧ da − h(a − c)ω1 ∧ ω2 = 0, ω2 ∧ dc = 0.

The last three equations are obtained by exterior differentiation of the three equations before them. Hence the differential system (170) is closed. To describe these surfaces observe that ω1 = 0,

(resp. ω2 = 0)

defines a family of lines of curvature, to be denoted by Γ2 (resp. Γ1 ). Along a curve of Γ2 , we have ω12 = 0, so that these curves are geodesics. Writing ω2 = ds, we have, along a curve of Γ2 , dx de2 de3 de1 = e2 , = ce3 , = −ce2 , = 0. ds ds ds ds Hence it is a plane curve with curvature c, the plane having the normal e1 . The last equation of (170) says that dc is a multiple of ω2 , which means that all the curves of Γ2 have the same Frenet equations and hence are congruent to each other. Since de1 = ω12 e2 + ω13 e3 = (he2 + ae3 )ω1 , the intersection of two neighboring planes of the curves of Γ2 is a line in the direction e1 × (he2 + ae3 ) = −ae2 + he3 . By (144) and (170), we have d(−e2 +

h h h e3 ) = + ω23 (−e2 + e3 ). a a a

Hence this direction is fixed. It follows that the planes of the lines of curvature in Γ2 are the tangent planes of a cylinder Z. The curves of Γ1 , being tangent to e1 , are the orthogonal trajectories of the tangent planes of Z. Each line of curvature of Γ1 is thus the locus of a point in a tangent plane of Z as the latter rolls about Z. The curves of Γ2 are the orthogonal trajectories of those of Γ1 . Each of them is therefore the position taken by a fixed curve on a tangent plane through the rolling. The surfaces defined by (170) can be kinematically described as follows: Take a cylinder Z and a curve C on one of its tangent planes. The surface M is the locus described by C as the tangent plane rolls about Z. Such a surface is called a

146


cylindrical molding surface. It depends on two arbitrary functions in one variable, one defining the base curve of Z and the other the plane curve C. On our molding surface the equation (156b) is completely integrable and has a solution t which depends on an arbitrary constant. We get in this way all non-trivial families of isometric surfaces preserving the lines of curvature. We observe that among the molding surfaces are the surfaces of revolution. The above discussion can be summarized as follows: In the three-dimensional euclidean space E 3 consider two pieces of surfaces M, M ∗ , such that: (a) their Gaussian curvature = 0 and they have no umbilics; (b) they are connected by an isometry f : M → M ∗ preserving the lines of curvature. Then M and M ∗ are in general congruent or symmetric. There are surfaces M , for which the corresponding M ∗ is distinct relative to rigid motions. The cylindrical molding surfaces, and only these, are such surfaces belonging to a continuous family of distinct surfaces, which are connected by isometries preserving the lines of curvature. In particular, we have proved Theorem 8.1. This is an example of an elaborate nature of a non-involutive differential system whose solutions are studied through successive prolongations. If the surface is considered as a map M → E 3 , then the first and second fundamental forms involve respectively the first and second order jets, h, k, m, n those of the third order, and p, q those of the fourth order. Hence the surface M must be of class C 5 for our proof to be valid. Problem B also leads to an over-determined differential system. Its treatment is more involved. We refer to Chern [1985] for details, and for a proof of Theorem 8.2.

Introduction

147

CHAPTER V THE CHARACTERISTIC VARIETY

In this chapter we will define the characteristic variety Ξ associated to a differential ideal (satisfying one non-essential restriction). This variety plays at least as important a role in the theory of differential systems as that played by the usual characteristic variety in classical P.D.E. theory. We shall give a number of examples of characteristic varieties, discuss some of their elementary properties, and shall state a number of remarkable theorems concerning characteristic varieties of involutive differential systems. The proofs of most of the results rely on certain commutative algebra properties of involutive tableaux and will be given in Chapter VIII. In Section 1 we define and give examples of the characteristic variety of a differential ideal having no Cauchy characteristics. Roughly speaking, it is given by all hyperplanes in n-dimensional integral elements whose extension fails to be unique. This is an infinitesimal analogue of saying that an initial value problem fails to have a unique solution, and as such is parallel to the classical meaning of characteristic. In Section 2 we define the characteristic variety of a linear Pfaffian differential system. This definition, which is the one we shall use throughout the remainder of the book, is modelled on the P.D.E. definition using the symbol. After showing that this symbol definition coincides with the previous one in the absence of Cauchy characteristics, we go ahead and explain in general the characteristics and the characteristic variety. Then we give a number of examples, including a local existence theorem for determined elliptic Pfaffian systems and the local isometric embedding of surfaces in E 3 . This latter example illustrates in a non-trivial manner essentially all the basic concepts—prolongation, involution, torsion, Cauchy characteristics, characteristic variety—in the theory of exterior differential systems. It will be carried as a running example in this chapter. In Section 3 we give some properties of the characteristic variety. The first few of these, such as the relation between the characteristic variety of a Pfaffian system and that of its prolongation, are elementary. Following this we turn to deeper properties, all of which require the use of the complex characteristic variety and require that the system be involutive. The first of these, Theorem 3.6, tells us “how many” local integral manifolds there are in terms of the dimension and degree of the complex characteristic variety. The second of these, Theorem 3.15, deals with the overdetermined case and relates the characteristic hyperplanes to the singular integral elements in dimension l where l is the character of the system (Definition 3.4). The third of these, Theorem 3.20, which is a differential system analogue of the theorem of Guillemin, Quillen and Sternberg [1970], states that the characteristic variety induces an involutive system in the cotangent bundle of integral manifolds. Here we shall only prove the result when the characteristic variety is smooth and consists only of isolated points, a case already found in Cartan (see Subsection (vi)), and a case that is simpler from a technical point of view. A number of examples illustrating this result will be given in Chapter VII. In this chapter, I will denote a differential ideal, i.e., a homogeneous ideal in

148

V. The Characteristic Variety

Ω∗ (M ) that is closed under exterior differentiation. Sometimes we shall refer to I as a differential system. We will denote by I ⊂ Ω1 (M ) the degree one piece of I, and we shall assume that I is given by the sections of a sub-bundle of T ∗ M that we also denote by I. When there is an independence condition it will be denoted by Ω, where Ω is a decomposable n-form defined up to non-zero scalar factors and up to adding n-forms from I; as in Chapter IV, Ω is given by a non-zero section of Λn (J/I) where J is a sub-bundle of T ∗ M with I ⊂ J ⊂ T ∗ M . For a submanifold N ⊂ M and differential form α ∈ Ω∗ (M ), we will denote the restriction α|N by αN . Thus a n-dimensional submanifold N ⊂ M is an integral manifold of the differential system with independence condition (I, Ω) if

αN = 0 for all α ∈ I . ΩN = 0

Finally, we will use the summation convention. §1. Definition of the Characteristic Variety of a Differential System. Let M be a manifold and Gn (T M ) → M the Grassmann bundle of n-planes in the tangent spaces to M . Points of Gn (T M ) will usually be denoted by (x, E) where x ∈ M and E ⊂ Tx M is an n-plane.1 Over Gn (T M ) we have the universal n-plane bundle U → Gn (T M ) whose fibre over (x, E) is just E. We shall consider the projectivization PU ∗ of the dual bundle U ∗ . A point in the fibre PUE∗ of PU ∗ over E ∈ Gn (T M ) will be written as [ξ] where ξ ∈ E ∗ \{0} is a non-zero vector and [ξ] ⊂ E ∗ is the corresponding line (the brackets are supposed to suggest homogeneous coordinates). By projective duality [ξ] determines a hyperplane [ξ]⊥ in E, and geometrically we may think of PUE∗ as being the set of hyperplanes in E ⊂ Tx M . Let I be a differential ideal and assume that I has no Cauchy characteristics, i.e., we assume that there are no vector fields v = 0 satisfying v

I ⊂ I.

This non-essential assumption is put here for convenience of exposition; it will be eliminated below. Let Gn (I) ⊂ Gn (T M ) be the set of n-dimensional integral elements of I. Associated to each hyperplane [ξ]⊥ ⊂ E is the polar space H(ξ) = {v ∈ Tx M : span{v, [ξ]⊥} is an integral element}, that we may think of as all ways of enlarging [ξ]⊥ to an n-dimensional integral element. Definition 1.1. The characteristic variety is the subset Ξ of PU ∗ defined by Ξ = ∪E∈Gn (I)ΞE 1 We

shall sometimes use the abbreviated notation E when the base point x ∈ M is unimportant.

§1. Definition of the Characteristic Variety of a Differential System

149

where ΞE = Ξ ∩ PUE∗ and ΞE = {[ξ] : H(ξ) E}. Thus, Ξ consists of all hyperplanes in n-dimensional integral elements whose extension to an n-dimensional integral element fails to be unique. To the extent that we think of integral elements as infinitesimal solutions to a differential system, the characteristic variety corresponds to non-uniqueness of an initial value problem, in close analogy to the classical notion. There is a commutative diagram of mappings Ξ ⊂ PU ∗ ↓ ↓ Gn (I) ⊂ Gn (T M ) and we shall denote by ΞE the fibre of Ξ → Gn (I) lying over E. The condition that [ξ] be characteristic is dim H(ξ) > n. Therefore, it does not depend on the particular E with [ξ]⊥ ⊂ E (so long as there is at least one such). Thus, we may give the following Definition 1.2. An (n − 1)-plane E n−1 ∈ Gn−1 (I) is non-characteristic in case dim H(E n−1 ) = n. From the proof of the Cartan–K¨ ahler theorem we have the result: Let I be a real-analytic differential system and N ⊂ M an (n − 1)-dimensional real-analytic integral manifold whose tangent planes are K-regular and non-characteristic. Then there is locally a unique extension of N to an n-dimensional integral manifold of I. It is easy to see that the fibre ΞE of the projection Ξ → Gn (I) is an algebraic subvariety of PE ∗ , i.e., it is defined by polynomial equations. This is because the polar equations are linear in vectors v ∈ Tx M , and ΞE consists of hyperplanes [ξ] for which the ranks of these equations jump suitably; this condition is expressed by homogeneous polynomial equations in ξ. Of importance will be the complex characteristic variety ΞC , defined as the complex solutions to these same polynomial equations. Equivalently, for a complex integral element E we may consider complex hyperplanes [ξ]⊥ in the complex vector space E, and then ΞC,E = {[ξ] ∈ PE ∗ : H(ξ) E} where the polar space H(ξ) = {v ∈ TC,x M : span{v, [ξ]⊥} is a complex integral element}.2 Of course, it may well happen that Ξ is empty but ΞC is not. The reason we assumed no Cauchy characteristics is that v ∈ H(ξ) for any Cauchy characteristic vector v. Thus, the characteristic variety should only be defined for integral elements that contain all Cauchy characteristic vectors. Equivalently, we may consider the differential system obtained by “foliating out” the 2 An n-plane E ⊂ T x, M is a complex integral element if αE = 0 for all α ∈ I; E need not be the complexification of a real integral element.

150


Cauchy characteristics and define the characteristic variety on this reduced system. For linear Pfaffian systems this annoyance will be circumvented. Before turning to more substantive examples, we mention these two: i) A Frobenius system when I is generated by dy1 , . . . , dys in Rn+s with coordinates (x1 , . . . , xn , y1 , . . . , ys ). Then ΞC = ∅. A converse of this for involutive systems will be discussed below. i i 2n ii) A Darboux system when I is generated by Θ = with i dx ∧ dy in R 1 n 1 n ∗ coordinates (x , . . . , x , y , . . . , y ). In this case, ΞE = PE is everything. Example 1.3 (Triply orthogonal systems, cf. DeTurck and Yang [1984]). Let X be a 3-dimensional Riemannian manifold and consider the following Problem. Determine the triples of foliations in X that intersect pairwise orthogonally. This problem was discussed in n-dimensions and in the real analytic case in Chapter III. Here we shall restrict to 3-dimensions and shall discuss the characteristic variety, which is the first step towards a C ∞ result. To set the problem up we let M 6 = F (X) be the bundle of orthonormal frames, on which we have the canonical parallelism given by the 1-forms ω1 , ω2 , ω3 , ω12 , ω13, ω23 satisfying the usual structure equations (here 1 ≤ i, j ≤ 3) (1)

dωi =

ωj ∧ ωji ,

ωij + ωji = 0,

j

that uniquely determine the connection matrix ωij . On M we let I be the differential ideal generated by the 3-forms Θi = ωi ∧ dωi As explained in Chapter III, the solutions to our problem are given by sections s : X → F(X) = M satisfying s∗ Θi = 0. Locally then we look for integral manifolds N 3 ⊂ M of I such that ΩN = 0 where Ω = ω1 ∧ ω2 ∧ ω3 . We shall denote p-planes in T M by E p (i.e., we don’t worry about the foot of E p , which is the point x ∈ M such that E p ⊂ Tx M ). When p = n we shall generally just write E. In case E p is an integral element of I we set r(E p ) = dim H(E p ) − p − 1. Thus, r(E p ) = 0 is the condition that E p extend to a unique integral E p+1 . Using (1), it is a nice little exercise to show that I is algebraically generated by the forms ω1 ∧ ω2 ∧ ω12 (2)

ω1 ∧ ω3 ∧ ω13 ω2 ∧ ω3 ∧ ω23 .


151

An integral E 3 on which Ω = 0 is therefore given by linear equations ⎛

⎞ ⎛ ω23 0 ⎝ ω13 ⎠ = ⎝ a21 ω12 a31

(3)

a12 0 a32

⎞ ⎛ ⎞ a13 ω1 a23 ⎠ ⎝ ω2 ⎠ = 0 0 ω3

where the aij are arbitrary. Thus, over each point of M the integral 3-planes on which Ω = 0 form an R6 . For E given by (3) with basis for E ∗ given by the restrictions (ω1 )E , (ω2 )E , (ω3 )E , a point ξ = [ξ1 , ξ2 , ξ3 ] ∈ PE ∗ corresponds to the hyperplane [ξ]⊥ ⊂ E defined by the additional equation ξ1 ω1 + ξ2 ω2 + ξ3 ω3 = 0.

(4)

Lemma 1.4. Setting r(ξ) = r([ξ]⊥), we have ⎧ ⎪ ⎨0 r(ξ) = 1 ⎪ ⎩ 2

if all ξi = 0 if one ξi = 0 if two ξi = 0.

Proof. We shall check the results when all the aij = 0; the general case is similar. By symmetry we may assume that ξ1 = 0 in (4); multiplying by a constant we may then assume that (4) is ω1 − αω2 − βω3 = 0.

(5)

Letting {ei , eij } be the dual frame to {ωi , ωij }, the integral E 2 given by ωij = 0 and (5) has basis v1 = αe1 + e2 v2 = βe1 + e3 . For v = (6)

λi ei +

i<j

λij eij we want to count the number of solutions of Θi , v1 ∧ v2 ∧ v = 0

with the transversality condition Ω, v1 ∧ v2 ∧ v = 0. By subtracting from v a linear combination of v1 , v2 and multiplying by a constant we may assume that v = e3 +

λij eij .

i<j

Then

v1 ∧ v2 ∧ v = e1 ∧ e2 ∧ e3 + e1 ∧ e2 ∧ ( λij eij ) i<j

−αe2 ∧ e3 ∧ ( λij eij ) + βe1 ∧ e3 ∧ ( λij eij ). i<j

i<j

152


Using (2) the equations (6) are λ12 = 0 αλ23 = 0 βλ13 = 0, from which the lemma (in the case aij = 0) is clear.

Since a hyperplane is characteristic exactly when r(ξ) > 0, we see that ΞE consists of the union of the three coordinate lines in PE ∗ ∼ = P2 ; i.e., it is the usual coordinate triangle encountered in plane projective geometry. The singular points of ΞE are clearly just the vertices of this coordinate triangle. If N ⊂ M is a local integral manifold, then N may be identified with X together with a framing, and a surface N 2 ⊂ N is characteristic if one leg of the framing is tangent to N 2 ; it is doubly characteristic (i.e., has tangent planes given by an intersection point of two branches in Ξ) if two legs of the framing are tangent to N 2 , which is equivalent to one leg being normal to N 2 . Example 1.5 (Linear Weingarten surfaces). Let X be an oriented Riemannian 3manifold and set M 5 = {(x, e3 ) ∈ T X : e3 = 1}. Thinking of e3 ∈ Tx X as corresponding to the oriented 2-plane e⊥ 3 ⊂ Tx X we may picture M as the manifold of oriented contact elements lying over X. We shall use the fibering picture π F (X) M ↓ X

π(x, e1 , e2 , e3 ) = (x, e3 ),

and for computational purposes shall pull all forms back to F (X). Among the forms that are well-defined on M are ω3 = e3 · dx dω3 = ω1 ∧ ω13 + ω2 ∧ ω23 (7)

Ω0 = ω 1 ∧ ω 2 Ω1 = ω1 ∧ ω23 − ω2 ∧ ω13 Ω2 = ω13 ∧ ω23 .

If N 2 ⊂ X is any oriented surface we have its canonical lift (Gauss map) γ:N →M where γ(y) = unit normal to N at the point y. The pull-backs of the forms (7) are γ ∗ (ω3 ) = 0 = γ ∗ (dω3 ) γ ∗ (Ω0 ) = induced area form dA γ ∗ (Ω1 ) = (Trace II)dA γ ∗ (Ω2 ) = (det II)dA


153

where II is the 2nd fundamental form of N ⊂ X. Conversely, a smooth surface γ : N 2 → M is the canonical lift of an immersed oriented surface in X if (i) γ ∗ (ω3 ) = 0 (ii) γ ∗ (Ω0 ) = 0. If only (i) is satisfied, then we shall speak of γ : N → M as being a generalized surface; these are special cases of Legendre manifolds (see the following remark). Note that ω3 ∧ (dω3 )2 = 0 on M , so that the differential ideal generated by ω3 has the Pfaff–Darboux local normal form. Remark. Given a differential ideal on a (2n + 1)-dimensional manifold generated by a 1-form satisfying θ ∧ (dθ)n = 0, the Pfaff–Darboux local normal form shows that the maximal integral manifolds N , called Legendre manifolds, are of dimension n and are given by one arbitrary function of n-variables. Here we may think of θ = dz − i yi dxi and N given by (xi , yi = ∂z(x) ∂xi ) where z = z(x) is an arbitrary function. In particular, the generalized surfaces are described locally by 1 arbitrary function of 2 variables. Definition 1.6. Let A, B, C be constants not all zero. The differential ideal I = {ω3 , Θ = AΩ2 + BΩ1 + CΩ0 } generated by ω3 and Θ will be called a linear Weingarten system. Remark that by the structure equations (1) together with dωij + ωik ∧ ωkj = 1 k l R 2 ijklω ∧ ω , it is easy to see that I is generated algebraically by ω3 , dω3 , Θ. The study of the integral elements thus leads us to the following linear algebra data: T is 5-dimensional vector space ω ∈ T ∗ is a 1-form

(= Tx M ) (= ω3 ∈ Tx∗ M )

Θ1 , Θ2 ∈ Λ2 T ∗ are 2-forms

(= dω3 , Θ ∈ Λ2 Tx∗ M ).

The integral 1-planes are E 1 = Rv1 are (i) (ii) (iii)

where ω, v1 = 0. Given E 1 its polar equations ω, v = 0 v1 Θ1 , v = 0 v1 Θ2 , v = 0.

In general the rank of these equations is 3, i.e., r(E 1 ) = 0. Geometrically, we expect (locally and in the real-analytic case) to be able to find a unique linear Weingarten surface of type (A, B, C) passing through a general curve Γ ⊂ X.

154


To compute the characteristics, we observe that these occur when equations (i)– (iii) become dependent. More precisely, for an integral 2-plane E ∼ = R2 we have ∗ ∼ 1 ∗ PE = P , and each [v1 ] ∈ PE such that the equations (i)–(iii) are dependent is characteristic. Since Θ1 and Θ2 do not contain ω, this is equivalent to equations (ii) and (iii) becoming dependent, i.e., when (8)

(λ1 Θ1 + λ2 Θ2 ) = 0

v1

for some (λ1 , λ2 ) = (0, 0). Note that if we restrict to integral elements E 2 on which Ω0 = 0 we must have λ2 = 0. Now it is easy to see that (8) is equivalent to (λ1 Θ1 + λ2 Θ2 )2 = 0. Setting λ = λ1 /λ2 it follows from (7) that this is the same as λ2 + B 2 − AC = 0.

(9)

Let λ± be the two roots of this equation (we allow complex values). Since dω3 and Θ are linearly independent, the two roots of (9) are the values for which the 2-form λdω3 + Θ = 0 becomes decomposable in the space of complex valued differential forms. We assume that λ+ = λ− and write λ+ dω3 + Θ = α+ ∧ β+ λ− dω3 + Θ = α− ∧ β− . Then it follows that ω3 , α+, β+ , α−, β− are pointwise linearly independent over C and that I is generated algebraically over C by ω3 , α+ ∧ β+ , α− ∧ β− . Proof. Since λ+ = λ− , we have 0 = ω3 ∧ (dω3 )2 =

1 (ω3 ∧ α+ ∧ β+ ∧ α− ∧ β− ). λ+ − λ−

Thus, over the complex numbers span{α+ ∧ β+ , α− ∧ β− } = span{dω3 , Θ}, which proves our claim that I = {ω3 , α+ ∧ β+ , α− ∧ β− }. Now, for any integral 2-plane E α+ ∧ β+ |E = 0 = α− ∧ β− |E . Moreover, we do not have α+ |E = β+ |E = 0 nor α− |E = β− |E = 0, since for example if the former holds then E is defined by ω3 = α+ = β+ = 0 and consequently α− ∧ β− |E = 0. Thus the restrictions to E of each of α+ , β+ and α− , β− spans a line in E ∗ , and these two lines are the characteristics. That is ΞE = [v+ ] ∪ [v−] where v± = 0 and α±(v± ) = β± (v± ) = 0.

§2. The Characteristic Variety for Linear Pfaffian Systems; Examples

155

§2. The Characteristic Variety for Linear Pfaffian Systems; Examples. In Chapter IV we have defined the class of linear Pfaffian systems. Associated to such a system are its tableau and symbol, and we shall now show how to compute the characteristic variety in terms of the symbol by a process that is formally analogous to that for P.D.E.’s. We begin by recalling some notation. From Chapter IV we recall that linear Pfaffian systems are given by sub-bundles I ⊂ J ⊂ T ∗M satisfying (10)

dI ⊂ {J}

where {J} ⊂ Ω∗ (M ) is the algebraic ideal generated by the sections of J.3 We set L = J/I, so that the exterior derivative induces a bundle mapping (11)

δ : I → (T ∗ M/J) ⊗ L

given in terms of bases by equation (56) in Chapter IV. Dualizing and using that (T ∗ M/J)∗ ∼ = J ⊥ ⊂ T M , giving δ is equivalent to giving the tableau mapping (cf. equation (60) in Chapter IV): (12)

π : J ⊥ → I ∗ ⊗ L.

The relations on the image of π are given by setting Q = I ∗ ⊗ L/image π and defining the symbol mapping σ to be the quotient mapping σ : I ∗ ⊗ L → Q. Then image π = kernel σ. æ We now define the characteristic variety Ξ ⊂ PL of a linear Pfaffian system. For 0 = ξ ⊂ Lx we let [ξ] ⊂ PLx be the corresponding line and define σξ : Ix∗ → Qx by σξ (w) = σ(w ⊗ ξ). Definition 2.1. The characteristic variety Ξ ⊂ PL is given by Ξ = ∪x∈M Ξx where Ξx = {[ξ] ∈ PLx : σξ fails to be injective}. 3 As usual, we shall use I and J to denote both a sub-bundle of T ∗ M and the C ∞ sections of that bundle—the context will make clear which use is intended.

156


Equivalently, Ξx = {[ξ] ∈ PLx : there exists 0 = w ∈ Ix∗ such that w ⊗ ξ ∈ Ax } where Ax is the tableau of (I, Ω) at x. We will now see how this definition looks in coordinates. Following the notations in §5 of Chapter IV, locally we choose 1-forms θa , ωi so that I = span{θa } J = span{θa , ωi } and then L∼ = span{ωi }. Here “span” means all linear combinations with smooth functions as coefficients. The condition (10) that the Pfaffian system be linear is dθa ≡ 0 mod {J} where {θa , ωi } = {J} is the algebraic ideal generated by the θa and ωi . This means that we have dθa ≡ πia ∧ ωi mod {I} where {I} = {θa } is the algebraic ideal generated by the θa ’s, and the πia are then 1-forms well-defined modulo J, and they give the tableau mapping (12) in coordinates. We may thus think of π as given by the tableau matrix ⎡

⎤ πn1 .. ⎥ . ⎦ . . . πns0

π11 . . . ⎢ .. π = ⎣. π1s0

mod J.

A basis for the relations on the image of π will be written as Baλi πia ≡ 0 mod J. Summarizing, the structure equations of a linear Pfaffian system are

(13)

(i) θa = 0 (ii) dθa ≡ πia ∧ ωi mod {I} (iii) Baλi πia ≡ 0 mod J (iv) Ω = ω1 ∧ · · · ∧ ωn = 0.

The symbol mapping σ is given in coordinates by the Baλi . More precisely, for ξ = ξi ωi (x) ∈ Lx as above, σξ is given by the matrix σξ = Baλi (x)ξi ∈ Hom(Ix∗ , Qx).


157

Then Ξx = {[ξ] : Baλi (x)ξi w a = 0 for some w = 0} = {[ξ] : rank Baλi (x)ξi < s0 } where rank I = s0 . It is clear that Ξx is defined by homogeneous polynomial equations in the ξi , so that Ξ = ∪x∈M Ξx is a family of algebraic varieties parameterized by M . The way to remember this definition is as follows: Associated to a P.D.E. system F λ (yi , z a , ∂z a /∂yi ) = 0 is the linear Pfaffian differential system θa = dz a − pai dyi = 0 dθa ≡ −dpai ∧ dyi mod {I} Ω = dy1 ∧ · · · ∧ dyn = 0 on the manifold M = {(yi , z a, pai ) : F λ (yi , z a, pai ) = 0}. Differentiation of the defining equations of M gives the symbol relations ∂F λ a dpi ≡ 0 mod J ∂pai where J = span{θa , dyi }. Comparing with equation (iii) in (13), we find that our definition of symbol relations for a linear Pfaffian differential system is a natural extension of the usual definition for a P.D.E. system. Correspondingly, our definition of the characteristic variety is the natural extension of the usual one for a P.D.E. system. We now want to compare the more general definition in the preceeding section with this one. Recall that the Cauchy characteristics are the vector fields v satisfying v v

θa = 0 dθa = 0 mod I.

If the 1-forms θa , ωi , πia fail to span Tx∗ M for some x, then by our constant rank assumptions this will be true in a neighborhood and we can find a vector field v satisfying v θa = v ωi = v πia = 0. By the structure equations (13) this vector field will be a Cauchy characteristic. Thus, under the assumption of no Cauchy characteristics we have (14)

span{θa (x), ωi (x), πia (x)} = Tx∗ M.

158


For any integral element (x, E) of (I, Ω), the independence condition ΩE = 0 implies that the restriction mapping Lx → E ∗ is an isomorphism. We will show that: Under this isomorphism, Ξx ⊂ PLx corresponds to ΞE ⊂ PE ∗ . In particular, in this case the characteristic variety ΞE depends only on the point x and not on the particular E lying over x. Proof. We omit reference to the point x. By (14), the integral element E will be given by linear equations πia − paij ωj = 0, where paij = paji by (ii) in (13). By a substitution πia → πia − paij ωj we may assume that E is given by (15)

θa = 0 πia = 0.

Additionally, by (iii) in Proposition 5.15 in Chapter IV we may assume that the symbol relations are given by Baλi πia ≡ 0 mod I.

(16) Finally, we may assume that

ξ = ωn . The polar equations of [ξ]⊥ are then easily seen to be (17)

θa = 0 a = 0. π1a = · · · = πn−1

It follows that E = H(ξ) if, and only if, the equations (15) are not consequences of the equations (17), equivalently, (18)

some πnb is not a linear combination of the 1-forms a }. {θa , π1a , . . . , πn−1

We choose a vector v such that some v πnb = 0 but where v θa = 0 and v πia = 0 for i = 1, . . . , n − 1 and all a. Contracting the symbol relations with v gives Baλn (v

πna ) = 0,


159

so that Baλn is not injective. Conversely, suppose that Baλn is not injective. We may assume that the kernel contains the vector (1, 0, . . . , 0), i.e., that all B1λn = 0. But then πn1 does not appear in any of the symbol relations (13), (iii); in particular, it is not a linear combinaa tion of the {θa , π1a , . . . , πn−1 }. Thus, the condition that ξ = ωn be characteristic according to either Definition 1.1 or Definition 2.1 is equivalent to (18). There is a very simple relation between the Cauchy characteristics and characteristic variety of a linear Pfaffian differential systems. Let A(I) ⊂ T M denote the Cauchy characteristic sub-bundle. Since A(I) ⊂ I ⊥ the mapping A(I) → L∗ given in coordinates by v → (ω1 (v), . . . , ωn (v)) is well-defined, and we denote its image by S ⊂ L∗ . Then S⊥ ⊂ L and we shall show that: The characteristic variety Ξ ⊂ PS ⊥ .

(19)

In particular, if S = 0 then it follows that the fibres Ξx of the characteristic variety lie in the proper linear subspaces PSx⊥ ⊂ PLx . Clearly, (19) also remains valid when we complexify. In the involutive case, there is a converse to the complex version of (19). Proof. Choose a basis ω1 , . . . , ωn for L so that ω1 , . . . , ωk is a basis for S ⊥ . Then for k + 1 ≤ ρ ≤ n there is a Cauchy characteristic vector field vρ satisfying vρ

k + 1 ≤ ρ, σ ≤ n.

ωσ = δρσ ,

From dθa ≡ 0 mod I

vρ we infer that

πρa ≡ 0 mod J. Thus the tableau matrix looks like ⎡ 1 π1 . . . πk1 ⎢. (20) ⎣ .. π1s0

...

πks0

0 ... .. .

⎤ 0 .. ⎥ .⎦

mod J.

0 ... 0

In particular, among the symbol relations we have πρa ≡ 0 mod J,

k + 1 ≤ ρ ≤ n.

160


From this it is easy to see that a characteristic vector ξ = ξi ωi must have ξk+1 = · · · = ξn = 0. We remark that whenever we may choose bases so that the tableau matrix has the block form (20), then the zeroes correspond to the image S ⊂ L∗ of Cauchy characteristics as described above. We also remark that the mapping A(I) → S may not be injective; using (ii) in (13) it is easy to see that this is the case exactly when the θa , ωi , πia fail to span T ∗ M . Examples of this arise by adding extra variables to any Pfaffian differential system. Example 2.2. We shall set up a linear Pfaffian differential system whose integral manifolds are the Darboux framings of immersed surfaces S ⊂ E 3 . For this we denote by F (E 3 ) → E 3 to the bundle of orthonormal frames (x, e1 , e2 , e3 ) in Euclidean 3-space. On F (E 3 ) we have the equations of a moving frame (here 1 ≤ i, j ≤ 3) (i) dx = ωi ei i

(21) (ii) dei =

ωij ej ,

ωij + ωji = 0

j

and structure equations (i) dωi =

ωj ∧ ωji

j

(22) (ii) dωij =

ωik ∧ ωkj .

k

We consider the Pfaffian differential system on M = F (E 3 ) (23)

(i) ω3 = 0 (ii) ω1 ∧ ω2 ∧ ω12 = 0.

An integral manifold of this system is given by an immersion f in the diagram f

(24)

N − → F (E 3 ) xf ↓ x E3

where dim N = 3, f ∗ (ω3 ) = 0, and f ∗ (ω1 ∧ ω2 ∧ ω12 ) = 0. From f ∗ (ω3 ) = 0 we have dxf = f ∗ (ω1 )e1 + f ∗ (ω2 )e2 , and it follows first that the image xf (N ) = S is an immersed surface in E 3 , and secondly that f(N ) consists of all Darboux frames (xf (y), e1 , e2 , e3 ) to S; here y ∈ N and e1 , e2 ∈ Txf (y) (S). The structure equations of the Pfaffian differential system (23) are (25)

dω3 = −ω13 ∧ ω1 − ω23 ∧ ω2


161

from which it follows that (23) is a linear Pfaffian differential system. The tableau matrix is ω13 ω23 0. To compute the Cauchy characteristic system of the Pfaffian differential system (23) we denote by ∂/∂ω1 , ∂/∂ω2 , . . . the tangent frame dual to the coframe ω1 , ω2 , . . . on F (E 3 ). Then using (25) it is easy to verify that: the Cauchy characteristic system of (23) is spanned by ∂/∂ω12 . This is an example of the block form (20). We now claim that ∂/∂ω12 is tangent to any integral manifold of (23). For this it will suffice to show that ∂/∂ω12 lies in any integral element of the system. Recall that an integral element is given by a 3-plane E ⊂ T F (E 3 ) satisfying the conditions (ω3 )E = 0, (dω3 )E = 0 (ω1 ∧ ω2 ∧ ω12 )E = 0. It follows after a short computation that E is given by linear equations ω13 − aω1 − bω2 = 0 ω23 − bω1 − cω2 = 0, and therefore ∂/∂ω12 ∈ E as claimed. The conclusion concerning integral manifolds (24) that we may draw is this: The fibering xf : N → S is given by the integral curves of the Cauchy characteristic foliation. Moreover, the equality Ξ = PS ⊥ holds in (19) above. Geometrically, the Cauchy characteristic curves correspond to spinning the tangent frame of S. The reason that Ξ equals all of PS ⊥ is that no conditions are being put on the immersed surfaces xf (N ) ⊂ E 3 ; thus we do not expect to be able to uniquely determine integral manifolds by data along a curve. We note that the pullbacks to N of the 1st and 2nd fundamental forms of S are given by the quadratic differential forms I = (ω1 )2 + (ω2 )2 II = a(ω1 )2 + b(ω1 ω2 ) + c(ω2 )2 . Remark finally that this discussion generalizes to Darboux framings associated to submanifolds Y n ⊂ X N where X is any Riemannian manifold and n, N are arbitrary. The Cauchy characteristics give the spinning of the tangential and normal frames to Y . We shall now give two further types of examples of characteristic varieties. For the first, following standard terminology we give the following Definition 2.3. The linear Pfaffian differential system (I, Ω) is said to be elliptic in case its real characteristic variety is empty Ξ = ∅. Example 2.4. In R2m = Cm we consider the Cauchy–Riemann system

(26)

∂u ∂v − i =0 i ∂x ∂y ∂u ∂v + =0 ∂yi ∂xi

i = 1, . . . , m.

162


As previously remarked, the symbol matrix of the Pfaffian differential system corresponding to (26) is given by its symbol matrix as a partial differential equation system, i.e., by the matrix ξ1 ξ2 ξ3 ξ4 . . . ξ2m−1 ξ2m (27) . −ξ2 ξ1 −ξ4 ξ3 . . . −ξ2m ξ2m−1 The real characteristic variety is, as expected, empty. However, for each z ∈ Cm the complex characteristic variety ΞC,z is given by the vanishing of all 2 × 2 minors of (27). It is easy to then verify that (28) where

ΞC,z = CP+m−1 ∪ CP−m−1 ⊂ CP 2m−1 √ √ CP±m−1 = {ξ2 = ± −1 ξ1 , . . . , ξ2m = ± −1 ξ2m−1 }.

For example, when m = 2 we may picture ΞC,z as two purely imaginary and conjugate skew lines in CP 3 . Example 2.5. We now consider a linear Pfaffian differential system (I, Ω) with square symbol matrices (in this case we say that the system is determined). Following the notation in (23), (iii) above we write the symbol relations of a determined system as Babi πia ≡ 0 mod J, 4 so that the symbol matrix is σξ = Babi ξi . We shall show that the system is involutive if for some ξ det σξ ≡ 0; i.e., if the complex characteristic variety is not everything, then the system is involutive. Proof. We may assume that det Babn = 0, and then by a basis change in the space of relations that Babn = −δab . When this is done the symbol relations are πna ≡ Bbaρ πρb

mod J

where 1 ≤ ρ, σ ≤ n − 1. It follows that the characters of the tableau matrix are given by s1 = s0 , . . . , sn−1 = s0 , sn = 0. Now write the symbol relations out as πna ≡ Bbaρ πρb + Bia ωi

mod I.

4 This notation is slightly misleading, since the symbol matrices σ = B bi ξ are elements of ξ a i Hom(W1 , W2 ) where W1 , W2 are different vector spaces of the same dimension. Thus we may pre- and post-multiply σξ by different invertible matrices.


163

Integral elements are given by linear equations θa = 0 πia − paij ωj = 0 where

(i) paij = paji (ii) panj = Bbaρ pbρj + Bja .

Choose paρσ = paσρ arbitrarily and use (ii) for 1 ≤ j ≤ n − 1 to determine the panσ = paσn . Then use (ii) when j = n to determine pann . It follows that the s0 n(n − 1)/2 quantities paρσ may be freely specified and that each set of these quantities uniquely determines an integral element; thus the space of integral elements is non-empty (the integrability conditions are satisfied) and has dimension equal to s0 n(n − 1)/2 = s1 + 2s2 + · · · + nsn . By Cartan’s test the system is involutive.

We will say that the Pfaffian differential system (I, Ω) is locally embeddable in case it is locally induced from the canonical system on J 1 (Rn , Rs0 ). It is easy to show that this is equivalent to J = span{θa , ωi } being a Frobenius system (cf. Proposition 5.10 in Chapter IV). Under this circumstance (I, Ω) is locally equivalent to a determined P.D.E. system (29)

F b (yi , z a, ∂z a /∂yi ) = 0,

and known results from P.D.E. theory may be applied to construct integral manifolds. We shall now prove that (30)

If the linear system (I, Ω) is determined and elliptic, and if n = rank(L) ≥ 4, then it is locally embeddable.

As a corollary of (30) we have the following result: (31)

Under the conditions of (30) there exist local integral manifolds of the Pfaffian system (I, Ω).

Proof of (31). We may locally realize (I, Ω) as the Pfaffian differential system associated to a determined 1st order elliptic P.D.E. system (29). Appealing to a standard result in elliptic equation theory (see Nirenberg [1973]), we infer that (29) has local solutions. Proof of (30). As noted above, the system (I, Ω) is involutive, and by the proof of that result we may write it as

(32)

(i) θa = 0 (ii) dθa = πia ∧ ωi + ηba ∧ θb (iii) πna = Bbaρ πρb mod J

164


where 1 ≤ ρ, σ ≤ n − 1 and where (33)

the 1-forms πρa are linearly independent

mod J = span{θb , ωi }.

The exterior derivative of (ii) in (32) gives 0 ≡ πia ∧ dωi

(34)

mod J.

We must show that, if n ≥ 3, this implies conversely that dωi ≡ 0

(35)

mod J.

For this we set a , ϕan = Bbaρ πρb , ϕa1 = π1a , . . . , ϕan−1 = πn−1

and using (iii) in (32) write (34) as ϕai ∧ dωi ≡ 0 mod J

(36) for each a.

Lemma 2.6. Assume that the system (I, Ω) is elliptic. Then (i) for each a the 1-forms ϕa1 , . . . , ϕan are linearly independent mod J; (ii) if n ≥ 3, then for each a = b the 1-forms {ϕai }, {ϕbi } are linearly independent mod J; and (iii) in general, for a1 < · · · < am and m ≤ n − 1 the 1-forms ϕa1 a , . . . , ϕan1 , . . . , ϕa1 m , . . . , ϕanm are linearly independent mod J. Proof. (i) Suppose there is a linear relation ζ ρ πρa + ζBbaσ πσb ≡ 0 mod J. Then ζ = 0 by (33); by homogeneity we may take ζ = −1 and then (δba ζ ρ − Bbaρ )πρb ≡ 0 mod J. By (33) this implies (37)

δba ζ ρ − Bbaρ = 0 for all ρ, b.

Taking 0 = ξ = (ξ1 , . . . , ξn ) with ξn = ξρ ζ ρ , (37) gives δba ξn − Bbap ξρ = 0. But this says that the ath row of the symbol matrix σξ is zero, which contradicts ellipticity. (ii) Given a linear relation (38)

(δca ζ ρ + ζBcaρ )πρc + (δcb η ρ + ηBcbρ )πρc ≡ 0 mod J


165

we choose ξ = 0 with ζ ρ ξρ = ξn ζ and η ρ ξρ = ξn η. This is possible since n ≥ 3, and (33) now gives as before ζ(δca ξn + Bcaρ ξρ ) + η(δcb ξn + Bcbρ ξρ ) = 0. But this implies that the ath and bth rows of the symbol matrix are dependent (it is easy to see that we don’t have ζ = η = 0); again this contradicts ellipticity. (iii) The proof just given applies to the general case, provided we can find a vector ξ = 0 that is orthogonal to m given vectors; under the conditions of the lemma, this is always possible. By (36), the above lemma and the usual Cartan lemma we have for each a (39)

dωi ≡ A(a)ij ∧ ϕaj

mod J

where the A(a)ij = A(a)ji are 1-forms. By ellipticity, the symbol gives a linear mapping Rn → s0 × s0 matrices, denoted by ξ → σξ , such that ξ = 0 ⇒ det σξ = 0. Thus s0 must be even, and since n ≥ 4 we cannot have s0 = 2, i.e., (40)

s0 ≥ 4.

We take a = b and use (39) for a and b to obtain (41)

A(a)ij ∧ ϕaj − A(b)ij ∧ ϕbj ≡ 0 mod J.

Again by Cartan’s lemma this implies that A(a)ij ∈ span{ϕak , ϕbk } mod J. Taking c = a, b and applying this also for a, c we infer that A(a)ij ∈ span{ϕak , ϕck } mod J. But then (40) and the lemma together imply that A(a)ij ∧ ϕaj ≡ 0 mod J which by (39) is our desired statement (35). Example 2.7. We shall study in some detail the isometric embedding system for an abstract Riemannian surface S mapping isometrically to E 3 . The notation S will denote the abstract surface, and S → S ⊂ E 3 will denote the isometric embedding. Although the study of this example for general dimensions and codimensions has been initiated in Chapter III and will be resumed in Chapter VII, we shall set it up somewhat differently. One motivation is that the characteristic variety will appear in a simple manner. Moreover, although we restrict here to the case of surfaces, all

166


aspects of the general theory already appear in this special case and the extension to higher dimensions basically involves more elaborate algebra. Finally, this example illustrates in a substantial way most of the aspects of the general theory of the characteristic variety. The general theory will be discussed further in Chapter VII below. Additional references are the original sources, Berger, Bryant and Griffiths [1983] and Bryant, Griffiths and Yang [1983], and the detailed exposition given in Griffiths and Jensen [1987] of these papers. In discussing this example, we shall make use of certain concepts such as prolongation, that was introduced in Chapter IV and will be discussed in Chapter VI below, and elementary results such as the relationship between the characteristic variety of a Pfaffian system and the characteristic variety of its 1st derived system, that also will be discussed below. These concepts and elementary results should be pretty much self-evident in our example. We begin by setting up the system and computing its structure equations and 1st prolongation. Geometrically the idea is to map the principal frame bundle of S to the Darboux frames of the image surface. In this regard, it may be helpful to keep in mind Example 2.2 above. We denote by π : P → S the principal frame bundle whose points are (y, e1 , e2 ) with y ∈ S and e1 , e2 an orthonormal basis of Ty (S) and where π(y, e1 , e2 ) = y. On P there is the canonical parallelism given by 1-forms ω1 , ω2 , ω12 satisfying the structure equations dω1 = −ω 2 ∧ ω 12 (42)

dω2 = ω1 ∧ ω12 dω12 = Kω 1 ∧ ω 2

where K is the Gaussian curvature of S. Now we set M = P × F(E 3 ) and on M consider the Pfaffian differential system given by (43)

(i) θi = ωi − ωi = 0, i = 1, 2, (ii) θ3 = ω3 = 0 (iii) ω1 ∧ ω2 ∧ ω12 = 0.

Throughout this example we shall use the index range i, j = 1, 2 and a = 1, 2, 3. We shall see below that the integrals of this system are locally graphs of maps f : P → F (E 3 ) where (i) f(P ) is the set F (S) of Darboux frames associated to an immersed surface S ⊂ E 3 and (ii) there is a commutative diagram f

(44)

P −−−−→ F (S) ⏐ ⏐ ⏐ ⏐π π" " xf

S −−−−→

S

where xf is an isometric immersion. One difference between this method and that in Chapter III is that here we make no a priori choice of frame field on S. Using the


167

equations (21) and (22) above of a moving frame and (42), the structure equations of (43) are

(45)

(i) dθ1 ≡ (ω12 − ω12 ) ∧ ω2 mod {θa } (ii) dθ2 ≡ −(ω12 − ω 12 ) ∧ ω1 mod {θa } (iii) dθ3 ≡ −ω13 ∧ ω1 − ω23 ∧ ω2 mod {θa }.

It follows that (43) is a linear Pfaffian differential system. Any 3-plane E ⊂ T (M ) on which all the 1-forms θa = 0 and ω1 ∧ ω2 ∧ ω12 = 0 is given by linear equations ω12 = p121 ω1 + p122 ω2 + p123ω12 (46)

ω13 = p11 ω1 + p12 ω2 + p1 ω12 ω23 = p21 ω1 + p22 ω2 + p2 ω12 .

Using (i) in (43) we may replace ωi by ωi in these equations. The conditions that (46) be an integral element is that all 2-forms dθa restrict to zero on E. By (45) this is p121 = p122 = 0, p123 = 1 p1 = p2 = 0 p12 = p21 . Setting p11 = a, p12 = p21 = b, p22 = c it follows that the space M (1) of integral elements of the system (43) is M × R3 , where (a, b, c) ∈ R3 and where the integral elements are given by ω12 − ω12 = 0 ω13 − aω1 − bω2 = 0 ω23 − bω1 − cω2 = 0. By definition, the 1st prolongation of (43) is the Pfaffian differential system on M (1) given by (i) θi = ωi − ωi = 0 (47)

(ii) θ3 = ω3 = 0 (iii) θ12 = ω12 − ω12 = 0 (iv) θ13 = ω13 − aω1 − bω2 = 0 (v) θ23 = ω23 − bω1 − cω2 = 0

together with the independence condition ω1 ∧ ω2 ∧ ω12 = 0.5 We note that the 1st prolongation (47) contains two parts: the original system (i) and (ii), and the equations (iii), (iv), (v) of integral elements of the original system. As discussed in Chapter IV, this is a completely general fact. We also note that equation (iii) is 5 Later

on, we shall see that M (1) should be taken to be M × (

Ê \{0}). 3

168


defined on M but did not appear in our original system, which was thus “incomplete”. More precisely, we will shortly see that this means that the original system fails to be involutive. Next we shall compute the Cauchy characteristic system of (43). For this we note first that the 1-forms appearing in (43) and (45) span all of T ∗ M , so that by our remarks following the proof of (19) above the mapping A(I) → S is injective. Setting θ12 = ω12 − ω12 , by (45) the tableau matrix of (43) is ⎤ ⎡ 0 −θ12 0 ⎣ θ12 0 0 ⎦ mod {θa , ωi , ωij }. (48) ω13 ω23 0 Referring to (20) above we see that there is, up to non-zero multiples, one Cauchy characteristic vector field. In fact, it is v = ∂/∂ω12 + ∂/∂ω12 . Geometrically it corresponds to spinning the tangent frames to S and S at the same rate. More precisely, as in example 3 above we may see that v lies in any integral element of (43), and therefore any integral manifold of this system is fibered by the circle group action whose infinitesimal generator is v. Using this observation it is easy to see that the integral manifolds of (43) are locally graphs of maps f : P → F (E 3 ) for which there is a commutative diagram (44) where xf is an isometric immersion. Now we observe that (43) fails to be involutive, essentially due to the fact that equation (iii) in (47) ω12 − ω12 = 0, which is implied by (i) and (ii) in (45) (uniqueness of the Levi–Civita connection), is missing. Referring to (48) the reduced characters are s1 = 2, s2 = 1, s3 = 0 ⇒s1 + 2s2 = 4, while as previously noted the space of integral elements is an R3 . Thus, Cartan’s test is not satisfied and (43) fails to be involutive. It is for this reason that we went ahead and wrote down the 1st prolongation (47) of (43). The next step is to compute the integrability condition and tableau for (47), and for this some additional notation will be helpful (we want to eliminate the indices—this is especially useful in the higher dimensional case). We introduce a vector space V ∼ = R2 and consider the following vector-valued differential forms ω = t (ω1 , ω2 ) and ω = t (ω 1 , ω2 ) are V -valued 1-forms $ $ # # ω 12 0 ω12 0 and ψ = are V ⊗ V ∗ -valued 1-forms ψ= −ω12 0 −ω 12 0 η = (ω13 , ω23) is a V ∗ -valued 1-form $ # 0 ω1 ∧ ω2 is a V ⊗ V ∗ -valued 2-form. ω∧ω = −ω1 ∧ ω2 0


169

ϕ = ω3 is a scalar 1-form. The structure equations of a moving frame now appear as (i) dω = −ψ ∧ ω + t η ∧ ϕ (ii) dψ = t η ∧ η (iii) dη + η ∧ ψ = 0.

(49)

Here, for example, ψ ∧ ω is the natural pairing (V ⊗ V ∗ -valued 1-form) ⊗ (V -valued 1-form) → V -valued 2-form. The structure equations (42) on the frame bundle P of S are (i) dω = −ψ ∧ ω

(50)

(ii) dψ = Kω ∧ ω

where ω ∧ ω is the obvious analogue of ω ∧ ω and K is the Gaussian curvature. The Pfaffian differential system (43) is now (i) θ = ω − ω = 0 (ii) ϕ = 0

(51)

together with the independence condition (iii) in (43). The structure equations (45) of (51) are (i) dθ ≡ −(ψ − ψ) ∧ ω mod {θ, ϕ} (ii) dϕ ≡ −η ∧ ω mod {θ, ϕ}.

(52)

The integral elements (46) for this system are given by ψ−ψ =0 η − Bω = 0 #

$ a b where B is the S V -valued function corresponding to the symmetric matrix . b c st (1) 3 2 ∗ The 1 prolongation (47) of (51) is given on M = P × F(E ) × S V by the Pfaffian differential system ⎧ (i) θ = ω − ω = 0 ⎪ ⎪ ⎪ ⎨ (ii) ϕ = 0 (53) ⎪ (iii) ψ − ψ = 0 ⎪ ⎪ ⎩ (iv) η − Bω = 0 2

∗

with the independence condition ω 1 ∧ ω 2 ∧ ω12 = 0. We want now to compute the integrability condition, tableau, and involutive prolongation of this system. For this we let γ : S2 V ∗ × S2 V ∗ → R

170


be the symmetric bilinear function obtained by polarizing the quadratic function γ(B, B) = det B = ac − b2 . Then letting ≡ denote congruence modulo the algebraic ideal generated by the system (53), we obtain from (49) and (50) that

(54)

(i) dθ ≡ 0 (ii) dϕ ≡ 0 (iii) d(ψ − ψ) ≡ −{γ(B, B) − K}ω ∧ ω (iv) d(η − Bω) ≡ −DB ∧ ω.

To explain equation (iv) we have d(η − Bω) = −η ∧ ψ − dB ∧ ω + Bψ ∧ ω − B t η ∧ ϕ (55)

≡ −Bω ∧ ψ − dB ∧ ω + Bψ ∧ ω modulo (53) = −DB ∧ ω

where DB is the S 2 V ∗ -valued 1-form defined by the coefficient of ∧ω in the right hand side of the middle equation in (55). From (54) we may draw two important conclusions: (56)

The integrability conditions of (53) are given by τ = γ(B, B) − K = 0.

Since this equation is not identically satisfied, the system is not involutive. (57)

The tableau of (53) is given by the S 2 V ∗ -valued 1-form DB.

We may picture it as the symmetric matrix # $ π11 π12 (58) π= π21 π22 where π11 ≡ da mod (53), π12 ≡ db mod (53), π22 ≡ dc mod (53). The symbol relations are given by (59)

π21 − π12 ≡ 0 mod (53).

Actually the tableau matrix of (54) should be ⎡ (60)

0 ⎢ 0 ⎢ ⎢ 0 ⎣ π11 π21

0 0 0 π12 π22

⎤ 0 0⎥ # 0 ⎥ 0⎥ = π ⎦ 0 0

$ 0 . 0

The top two rows of zeros correspond to the original system (43), which as reflected by (i) and (ii) in (54) has gone into the 1st derived system of (53). (As noted in §6 of Chapter IV, this also is a general property of prolongation.) The third row of


171

zeros corresponds to (iii) in (54), where we recall that the tableau matrix is always considered modulo J = span{θa , ωi } where the θa span the system (53). The last column of zeros reflects the Cauchy characteristic vector field. It is an easily verified fact that the involutivity of a tableau (60) is equivalent to the involutivity of the non-zero block (58). In other words, in testing for involutivity of a tableau we may throw out the first derived system and Cauchy characteristic system. According to the general prolongation scheme, as explained more fully in Chapter VI below, we must set the integrability conditions equal to zero. This gives the Gauss equations (61)

γ(B, B) = K.

It is easy to see that these equations always have solutions, and that the subset of M (1) defined by (61) and B = 0 is a smooth manifold Y . We therefore restrict the Pfaffian differential system (53) to Y ; in effect this means that we impose the condition (61) and the exterior derivative of this equation. To compute the latter we have the easy Lemma 2.8. dγ(B, B) = 2γ(B, DB) where DB is defined by (55). It follows that on Y we must add to (59) the additional symbol relation (here ≡ denotes congruence modulo the system (53)) (62)

γ(B, π) ≡ dK

where π is the S 2 V ∗ -valued 1-form 2DB. Definition 2.9. We shall call the restriction of (53) to Y , as defined by the Gauss equation (61) and B = 0, the involutive prolongation of the isometric embedding system (51). This terminology will be justified in a moment. Remark that the involutive prolongation is given by the system (53) on Y where the structure equations are now

(63)

(i) dθ ≡ 0 (ii) dϕ ≡ 0 (iii) d(ψ − ψ) ≡ 0 (iv) d(η − Bω) ≡ −π ∧ ω #

with π=

π11 π21

π12 π22

$

and where the symbol relations are (64)

(i) π12 − π21 ≡ 0 (ii) γ(B, π) ≡ dK.6

6 We recall our notation that ≡ denotes congruence modulo the algebraic ideal generated by the 1-forms in (53).

172


Summarizing, the procedure is this: (i) Begin with the naive isometric system (43) = (51). The tableau of this system is not involutive and so we must prolong. In effect, prolonging means that we equate the connection forms and add all candidates aω12 + 2bω1 ω2 + cω22 for the 2nd fundamental form as new variables. (ii) For the prolonged system (47) = (53) the integrability condition is given by the Gauss equations (61). So it also fails to be involutive, and we adjoin the Gauss equations together with their exterior derivatives. We shall now prove that the resulting system (63) is involutive. By the remarks above we may restrict our attention to the essential piece # $ π11 π12 π= π21 π22 of the tableau, whose symbol relations (64) are (i) π12 − π21 ≡ 0 (ii) aπ22 − bπ12 − bπ21 + cπ11 ≡ dK

(65) #

$ a b where B = . Remark that b c (66)

dK ≡ 0 mod {ω1 , ω2 }.

Since B = 0, by choosing a general basis for V we may assume that a = 0 (this will correspond to choosing a regular flag). It follows that π11 , π21, may be assigned arbitrarily, and then π12 is determined by (i) in (65) and π22 by (ii) in (65). Moreover, by (66) we may choose π22 to annihilate the integrability condition. The characters are s1 = 2, s2 = 0. Integral elements are given by linear equations πij − pijk ωk = 0 where, upon setting dK = Kj ωj , (i) pijk = pikj (ii) p12j = p21j (ii) ap22j − 2bp12j + cp11j = Kj . From (i) and (ii) it follows that pijk is symmetric in all indices, and from (iii) it follows that pijk is determined by p111 and p112 . Thus the integral elements lying over a point of Y have dimension equal to 2 = s1 + 2s2 , and so by Cartan’s test the system is involutive. Finally, we want to compute the characteristic variety of the involutive prolongation of the isometric embedding system. We have noted above that if the tableau matrix looks like # $ 0 π= , π ˜

§3. Properties of the Characteristic Variety

173

then as is easily verified the characteristic variety for π is the same as that for π ˜. Thus we must compute the characteristic variety Ξ for a tableau matrix #

π11 π21

π12 π22

0 0

$

with symbol relations (65) where we have absorbed dK into π22 . Recall from (19) above that for, y ∈ Y , the Cauchy characteristics as reflected in the last column of zeros, give Ξy ⊂ P1 ⊂ P2 . The determination of Ξy ⊂ P1 is consequently the same as that of determining the characteristic variety of the tableau matrix #

π11 π21

π12 π22

$

The symbol matrix at ξ = [ξ1 , ξ2 ] ∈ P1 is $ −ξ1 ξ2 . σξ = (cξ1 − bξ2 ) (aξ2 − bξ1 ) #

Then det σξ = a(ξ2 )2 − 2bξ1 ξ2 + c(ξ1 )2 is a quadratic form with discriminant ∆(σξ ) given by ∆(σξ ) = ac − b2 = K where K is the Gaussian curvature. Thus: (67)

If K(y) < 0 then Ξy ⊂ P2 consists of two distinct points. If K(y) = 0 then Ξy consists of one point counted twice. Finally, if K(y) > 0 then Ξy = ∅ but ΞC,y ⊂ CP 1 consists of a pair of distinct conjugate points.

Following the usual P.D.E. terminology we may say that in the cases K < 0, K = 0, K > 0 the involutive prolongation of the isometric embedding system is respectively hyperbolic, parabolic and elliptic. For a surface S ⊂ E 3 with K < 0, at each point p ∈ S the two characteristic lines in Tp (S) are the asymptotic directions. A striking fact is that the isometric embedding system for M n ⊂ E n(n+1)/2 , is never elliptic when n ≥ 3 (cf. Bryant, Griffiths and Yang [1983] and the references cited there).

174


§3. Properties of the Characteristic Variety. In this section we shall state a number of properties of the characteristic variety of a linear Pfaffian differential system (I, Ω). The proofs of the more substantial of these will be given in Chapter VIII. (i) The 1st derived system and Ξ. We consider a Pfaffian differential system given by a filtration I ⊂ J ⊂ T ∗ X and with 1st derived system (cf. Chapter I) I1 = ker{δ : I → Λ2 T ∗ M

mod {I}}.

For an adapted basis {θ1 , . . . , θp ; θp+1 , . . . , θs0 } = {θρ , θε } for I1 ⊂ I (here 1 ≤ ρ, σ ≤ p and p + 1 ≤ ε, δ ≤ s0 ) we have (i) dθρ ≡ 0 mod {I} (ii) dθε ≡ πiε ∧ ωi mod {I}. Here we recall that {I} is the algebraic ideal generated by the sections of I. The symbol relations are of the form (i) πiρ ≡ 0 mod J (ii) Bελi πiε ≡ 0 mod J, #

and the tableau matrix is

$ 0 . πiε

To put this in an intrinsic algebraic settting, a sub-bundle I1 ⊂ I gives a quotient dual bundle, i.e., we have 0 → (I/I1 )∗ → I ∗ → I1∗ → 0. The tableau corresponding to the above matrix is given by a sub-bundle A ⊂ I ∗ ⊗L with the property that A projects to zero in I1∗ ⊗ L, i.e., A ⊂ (I/I1 )∗ ⊗ L ⊂ I ∗ ⊗ L. For 0 = ξ ∈ L the symbol mapping σξ : I ∗ → Q restricts to

σ1,ξ : (I/I1 )∗ → Q,

and it follows directly from the definitions that ker σξ = ker σ1,ξ . Thus the characteristic variety is the same as if we consider only the bottom nonzero block in the tableau matrix, i.e., we consider only the “smaller” symbol matrices Bελi ξi .


175

Informally, we may rephrase this as: (68)

In computing the characteristic variety, we may ignore the 1st derived system.

This property of characteristic varieties may be viewed as a generalization of the fact that the characteristic variety of a P.D.E. system depends only on its highest order terms. As was discussed in §6 of Chapter IV, if we prolong a Pfaffian differential system the original system appears in the 1st derived system of its prolongation. We have already encountered this phenomenon in Example 2.7 above (compare (51), (53), and (54)), where property (68) was in fact used. (ii) 2nd order systems and Ξ. We begin with a Pfaffian differential system that “looks like” the system associated to a 2nd order P.D.E. system. Rather than giving an involved intrinsic formulation of this we shall use indices. Thus we assume the system to be given by θa = 0 θia = 0 (69)

dθa ≡ 0 mod {I} a ∧ ωj mod {I} dθia ≡ πij

where {I} = {θa , θia } and the symbol relations are (70)

a a (i) πij ≡ πji mod J a (ii) Baλij πij ≡ 0 mod J

and where in (ii) it is understood that Baλij = Baλji. Example 3.1. We consider a 2nd order P.D.E. system a ∂2 za λ i a ∂z y , z , i , i j = 0. (71) F ∂y ∂y ∂y To set this up as a Pfaffian differential system we use the space J 2 (Rn , Rs0 ) of 2jets of maps from Rn to Rs0 , and on J 2 (Rn , Rs0 ) we use coordinates (yi , z a, pai , paij ) where paij = paji . In J 2 (Rn , Rs0 ) we consider an open subset M of smooth points on the locus F λ (yi , z a , pai , paij ) = 0. Setting θa = dz a − pai dxi |M θia = dpai − paij dxj |M a πij = −dpaij |M

ωi = dxi |M

176


(this is just the restriction to M of the canonical system) and using dθa = −dpai ∧ dxi ≡ 0 mod {θia }, we see that the Pfaffian differential system corresponding to (71) has the form (69) with symbol relations (70) where Baλij =

∂F λ ∂pij a

.

The usual symbol matrix of the P.D.E. system (71) is given by Baλij ξi ξj ,

(72)

and we want to extend this to the Pfaffian differential system (69). More precisely, it is well known that the characteristic variety associated to the symbol matrix (72), in which ξ appears quadratically in each term, is the same as that obtained by writing (71) as a 1st order system and computing its symbol matrix, in which ξ appears linearly in each term. It is the differential system analogue of this that we wish to establish. We remark that the condition that (69) be locally induced from a 2nd order P.D.E. system, i.e., that there should locally be an embedding f : X → J 2 (Rn , Rs0 ) satisfying span{θa , ωi } = span{f ∗ dxa , f ∗ dyi } span{θia , θa , ωi } = span{f ∗ dpai , f ∗ dz a , f ∗ dyi }, is the Frobenius condition dθa ≡ 0 mod {θa , ωi } dωi ≡ 0 mod {θa , ωi } (these plus (69) imply that dθia ≡ 0 mod {θa , θia , ωi }). Example 3.2. The involutive prolongation (53) of the isometric embedding system for S in E 3 has the form (69) (where span{θa } = span{θ, ϕ} and span{θa , θia } = span{θ, ϕ, ψ − ψ}), so that it “looks like” a 2nd order P.D.E. system. (This is certainly natural to expect, since the curvature is involved.) But it follows from (49) and (50) that it is not locally equivalent to such a system. Returning to the general discussion, we want to determine the characteristic variety of the system (69). The tableau matrix has the block form # $ 0 π= , a πij and by (68) we may ignore the block of zeros. For ξ = ξi ωi , the symbol matrix applied to a vector w = {wia } is a wi ξj − wja ξi (73) σξ (w) = . Barij wia ξj


177

The two blocks of this column vector correspond to the blocks (i) and (ii) of symbol relations (70). If σξ (w) = 0, then from wia ξj = wja ξi ,

ξ = 0,

we conclude that wia = w a ξi is a decomposable tensor. Using this the second block in (73) gives (74)

Baλij w a ξi ξj = 0.

In other words, the symbol matrix (72) is singular. Conversely, if (74) holds, then for wia = w a ξi we have σξ (w) = 0. In conclusion: (75)

The characteristic variety for the Pfaffian differential system (69) and (70) is the same as the characteristic variety formed from the symbol matrices (72).

Informally, we may say that if the tableau matrix of a Pfaffian differential system looks like the tableau matrix of a 2nd order P.D.E. system, then the characteristic variety according to Definition 2.1 above may be computed as one ordinarily would for 2nd order P.D.E. systems. Of course, this may be generalized to higher order systems. Example 3.3. Referring to (64), the characteristic variety of the involutive prolongation of the isometric embedding system is given immediately by a(ξ2 )2 − 2bξ1 ξ2 + c(ξ1 )2 = 0, a result we arrived at there by a somewhat longer calculation following the original definition. (iii) Characteristic variety of the 1st prolongation. We have briefly introduced the 1st prolongation in Chapter IV and will more fully discuss it in Chapter VI below. Here, we shall show by computation how the characteristic variety behaves under prolongation. We consider a linear Pfaffian differential system whose integrability conditions are assumed to be satisfied; thus there are integral elements over each point. Omitting reference to the independence condition, such a system may be assumed to be given by (13) and (16) above (i) θa = 0 (ii) dθa ≡ πia ∧ ωi mod {I} (iii) Baλi πia ≡ 0 mod {I} where (iii) are a basis for the symbol relations. Writing these as modulo {I} = {θb } means that the torsion has been absorbed. Integral elements are given by linear equations πia − paij ωj = 0

178


where paij = paji Baλi paij = 0. The 1st prolongation is a differential system on the manifold M (1) obtained locally from M by adding the paij ’s satisfying these conditions. On M (1) the 1st prolongation is the Pfaffian differential system I (1) generated by the equations (i) θa = 0 (ii) θia = πia − paij ωj = 0

(76)

(iii) Baλi paij = 0. The exterior derivatives of these equations give, using the original structure equations, dθa ≡ 0 mod {I (1) }

(77)

a ∧ ωj mod {I (1) } dθia ≡ πij

where locally {I (1) } = {θa , θia } and a = −dpaij + (horizontal forms relative to M (1) → M ).7 πij

Comparing (76) and (77) we see as before that the original Pfaffian differential system goes into the 1st derived system of its prolongation, and hence, by (68), when computing the characteristic variety of I (1) we need only consider the tableau a πij . Differentiating equation (iii) in (76) and using that

horizontal forms relative to M (1) → M

= span{θa , ωi , πia}8 = span{θa , π i , θia }

where span allows linear combinations with coefficients in C ∞ (M (1)), it follows that a the symbol relations on the πij are (78)

a a πij ≡ πji mod {θb , θib , ωi } a ≡ 0 mod {θb , θib , ωi }. Baλi πij

Note that the 2nd set of relations is indexed by pairs (λ, j) of indices. Thus we should write these relations as a Ba(λ,j)ik πik ≡ 0 mod {θb , θib , ωi }

differential form ϕ ∈ Ωq X is horizontal relative to a smooth mapping f : X → Y if ϕ(x) ∈ π∗ ∧q Tf∗(x) Y for all x ∈ X. 8 This equality is valid if there are no Cauchy characteristics. The general case may be done by “foliating out” the Cauchy characteristics or by the more intrinsic argument used in §6 of Chapter IV to prove the same equations as (78) below—cf. equations (118) there. 7A


179

where Ba(λ,j)ik = δkj Baλi . If ξ = {ξi } is in the characteristic variety for (I (1) , Ω), then there exists w = {wia } satisfying (i) wia ξj = wja ξi (λ,j)ik

(ii) Ba

wia ξk = 0.

The reasoning here is now analogous to that used in establishing (75) just above. From (i) it follows that wia = w a ξi , and then from (ii) it follows that Baλi w a ξi ξj = 0,

1 ≤ j ≤ n.

It follows that Baλi w a ξi = 0, which says that ξ is characteristic for (I, Ω). Since the converse is obviously true, we have established that: (79)

Under the projection ω ˜ : M (1) → M , there is a natural isomorphism ∼ ˜ Ξ(1) x = Ξω(x) between the fibre of the characteristic variety Ξ(1) for (I (1), Ω) lying over x ∈ M (1) and the fibre of the characteristic variety Ξ for (I, Ω) lying over ω ˜ (x) ∈ M .

Informally, we may say that, in the absence of integrability conditions, the characteristic variety remains unchanged when we prolong. If there are integrability conditions, then they will contribute additional symbol relations to the prolonged (1) system and the characteristic variety may get smaller—i.e., Ξx may be a proper subvariety of Ξω(x) . ˜ Remark. The question of whether the symbol relations (78) may be refined to (78bis)

a a πij ≡ πji mod {I (1)} a ≡ 0 mod {I (1) }, Baλi πij

or equivalently whether the absence of integrability conditions on M implies the absence of integrability conditions on M (1) , is an interesting one. In general the answer is no; however, it will be proved in Chapter VI below that (1) (I, Ω) involutive (I , Ω) involutive ⇒ . on M on M (1) The proof will show that the involutivity of the tableau of (I, Ω) implies both that the tableau of (I (1), Ω) are involutive and that there are no integrability conditions for (I (1) , Ω), which is just (78bis). This argument will be put in a conceptual framework in Chapter VIII when we discuss Spencer cohomology.

180


(iv) Relationship between the characteristic variety and the Cartan– K¨ ahler theorem. The remaining properties of the characteristic variety are more substantial; they require involutivity and deal with the complex characteristic variety. Let (I, Ω) be a linear Pfaffian differential system on a manifold M given by a filtration I ⊂ J ⊂ T ∗ M . We give the structure equations of (I, Ω) in the form (13) above. For simplicity of exposition we assume that there are no Cauchy characteristics, so that span{θa , ωi , πia} = T ∗ M . The statements of the results given below remain valid without this assumption. From section 5 in Chapter IV we recall the tableau matrix given by equation (88) there

(80)

π=

π11 . . . .. .

πn1 .. .

π1s0

πns0

...

mod J,

(we now omit the bars over the πia ’s and understand that all 1-forms in π are considered modulo J), and the reduced Cartan characters s1 , s2 , . . . , sn defined inductively by number of linearly independent (81) s1 + · · · + sk = . forms in the first k-columns of π Here we assume that the basis ω1 , . . . , ωn for J/I is chosen generically. Also, in reality (81) is defined at each point x ∈ M , and we assume that these pointwise defined ranks are constant. We also recall that, in the absence of integrability conditions, the reduced characters are equal to the usual characters sk —cf. equation (86) in Chapter IV. Definition 3.4. The character l and Cartan integer κ of (I, Ω) are defined by

s1 , . . . , sl = 0, sl+1 = · · · = sn = 0 κ = sl .

As will be seen below, both the character and Cartan integer are invariant under prolongation so long as the system has no integrability conditions. Now suppose that the system is involutive and real analytic. According to the Cartan–K¨ ahler theorem we may construct local integral manifolds for (I, Ω) by solving a succession of initial value or Cauchy problems. Such a succession of initial value problems corresponds to nested sequence of integral manifolds 0 N ⊂ N 1 ⊂ · · · ⊂ N n−1 ⊂ N dim N p = p whose tangent spaces form a regular flag. From Chapter III we recall that: N is uniquely determined by Nl and Nl is uniquely obtained from Nl−1 by prescribing “κ arbitrary functions of l variables”. Thus we may think of (l, κ) as telling us something about how many local integral manifolds there are. To an algebraic geometer, l resembles a transcendence degree


181

and κ a field extension degree—this analogy will turn out to be precise. We will state results that express l, κ and the condition to be a regular flag in terms of algebro-geometric properties of the complex characteristic variety ΞC ⊂ PLC . In practice these theorems usually allow us to determine l, κ and regular flags without going through the sometimes laborious procedure of calculating the si . This will be illustrated by examples. Assume that (I, Ω) is involutive and let σ : IC∗ ⊗ LC → QC be the complexified symbol map for (I, Ω). For each x ∈ M and 0 = ξ ∈ LC,x we have ∗ σx : IC,x ⊗ LC,x → QC,x ∗ σx,ξ : IC,x → QC,x ∗ where for w ∈ IC,x

σx,ξ (w) = σx(w ⊗ ξ).

By definition ΞC,x = ([ξ] ∈ PLC,x : dim ker σx,ξ ≥ 1}. It is clear that ΞC,x is a complex algebraic variety, in fact, the ideal of ΞC,x is by definition the homogeneous ideal generated by suitable minors of the symbol matrix Baλi (x)ξi . Assuming first for simplicity that ΞC,x is irreducible we set d = dim ΞC,x δ = deg ΞC,x µ = dim ker σx,ξ where [ξ] ∈ ΞC,x is a general point. As a consequence of the involutivity of (I, Ω) we will see that d, δ, and µ are independent of x ∈ M . In general, if (α) ΞC,x = ΞC,x α

is the unique decomposition of ΞC,x into irreducible components we set (α)

d = max dim ΞC,x α

δ= µ(x) = (α)

δ (α) (x) µ(α) (x) (α)

where δ (α) (x) = deg ΞC,x , µ(α)(x) = dim ker σx,ξ (α) where [ξ (α)] ∈ ΞC,x is a general point, and the sum over components of maximal dimension. Again, d, (α) denotes µ (x)δ (α) (x) will be independent of x. δ, and

182


Example 3.5. For the involutive prolongation of the isometric embedding system we have ξ2 −ξ1 σx,ξ = cξ1 − bξ2 aξ2 − bξ1 where ac − b2 = K(x) is the Gaussian curvature. If K(x) = 0 then ΞC,x consists of two distinct points with each having δ (α) (x) = 1, µ(α) (x) = 1. If K(x) = 0 then ΞC,x consists of one point with δ = 2 and µ = 1. Note that (82)

µ(1) (x)δ (1)(x) + µ(2)(x)δ (2) (x) = 2, K(x) = 0 µ(x)δ(x) = 2, K(x) = 0

in accordance with the above remark. Indeed, if K(x) = 0 then over C we may assume that the 2nd fundamental form is a 0

0 , c

ac = K = 0,

in which case ξ2 −ξ1 cξ1 aξ2 % $ # % $ # c c = i, ∪ i, − . a a

σx,ξ = ΞC,x

For each point of ΞC,x clearly this matrix has rank one. If K(x) = 0 then we may assume that the 2nd fundamental form is a 0 , 0 0

a = 0.

In this case, ΞC,x = [1, 0] counted with multiplicity 2 and at this point σx,ξ =

0 −1 0 0

has rank one. By our assumption of involutivity the 2nd fundamental form can never be zero. Our result is the following Theorem 3.6. Let (I, Ω) be an involutive Pfaffian system of character l and having Cartan integer κ. Then l = d+1 (α) κ= µ (x)δ (α)(x).


183

Corollary 3.7. Suppose that all µ(α)(x) = 1 (this is frequently the case). Then l = dim ΞC,x + 1 κ = deg ΞC,x. If we omit reference to the particular point x ∈ X we may rephrase this as: (83)

The integral manifolds of an involutive, real analytic Pfaffian-differential system locally depend on deg ΞC arbitrary functions of dim ΞC + 1 variables.

Example 3.8. Referring to (82) above, the local isometric embeddings of a real analytic surface in E 3 depend on two arbitrary functions of one variable. We want to explain this, and the result we shall find is essentially: To locally isometrically embed S in E 3 , we choose a connected curve γ ⊂ S. Then the isometric embeddings γ → γ ⊂ E3 depend on two functions of one variable, and such an embedding extends essentially uniquely to a local isometric embedding S → S ⊂ E3 provided that γ is suitably general. We let (I, Ω) on the manifold M denote the involutive prolongation of the isometric embedding system. Then M ⊂ P × F(E 3 ) × S 2 V ∗ . There is an essentially unique lifting of γ ⊂ S to P given by s → (y(s), e 1 (s), e2 (s)) where γ is given by s → y(s) and e1 (s) is the unit tangent to γ (s is an arclength parameter). To be ‘essentially unique’ will mean that the lifting is unique once we have specified it at one point. The connection form ω12 = kg (s)ds, where kg is the geodesic curvature of γ in S. Let γ → γ be given by s → x(s) ∈ E 3 and set e1 (s) = dx(s)/ds, which is a unit vector. This embedding locally depends on two functions of one variable. We claim that there are essentially unique vectors e2 (s), e3 (s) and functions a(s), b(s) such that de1 = kg e2 + ae3 ds de2 (ii) = −kg e1 + be3 ds de3 (iii) = −ae1 − be2 . ds (i)

184

To see this we note that


& & & de1 &2 2 2 2 & & & ds & = kg + a = κ

where κ(s) is the curvature of γ in E 3 . Solving this gives a(s) up to ±1. Referring to (i) above, we may find unique unit vectors e2 (s) and e3 (s) in e1 (s)⊥ such that de1 /ds is the hypothenuse of the right triangle with sides kg e2 and ae3 . Having determined e2 we may determine b up to ±1 from the length of de2 /ds. Having determined a(s) and b(s), we may determine c(s) by the Gauss equation a(s)c(s) − b(s)2 = K(y(s)). In this way, given γ → γ we have determined a frame e1 (s), e2 (s), e3 (s) and 2nd fundamental form a(s)ω12 +2b(s)ω1 ω2 +c(s)ω22 along γ, i.e. we have a 1-dimensional integral manifold N 1 ⊂ M lying over the graph of γ → γ. Since s2 = 0 this integral manifold extends locally to a unique integral manifold N 2 of (I, Ω). æ Example 3.9. We consider a single linear P.D.E. (84)

P (x, D)u = 0

of order m with one unknown function. Here we use the standard notations Pα (x)Dxα P (x, D) = |α|≤m

Dxα

= (∂/∂x1 )α1 . . . (∂/∂xn )αn

|α| = α1 + · · · + αn . As follows from the discussion above, when we write (84) as a Pfaffian differential system and compute its characteristic variety we get the expected answer ΞC = {(x, ξ) : Pm (x, ξ) = Pα (x)ξ α = 0}. |α|=m

This system (84) also turns out to be involutive (this is a nice exercise using Cartan’s test). Since clearly also all µ(α)(x) = 1 (the symbol is a 1 ×1 matrix) it follows from (83) that, in the real analytic case, the solutions to (84) depend on m arbitrary functions of n − 1 variables. These are just the values of u and its 1st m − 1 normal derivatives along a non-characteristic hypersurface. In this case the result is a well known consequence of the Cauchy–Kowaleski theorem. Example 3.10. We reconsider the Cauchy–Riemann system given by (26) above with symbol matrix (27) and complex characteristic variety (28) there. Then δ (1) = δ (2) = µ(1) = µ(2) = 1, d = m−1 κ=2 and (83) is in accordance with the well-known fact that holomorphic functions in an open set U ⊂ Cm depend on two real functions of m real variables (think of locally extending a complex-valued real analytic function from Rm to Cm ). As another consequence of Theorem 3.6 we have the following:


185

Corollary 3.11. Let (I, Ω) be a C ∞ involutive linear Pfaffian differential system whose complex characteristic variety is empty, i.e., ΞC = ∅. Then I is completely integrable. In particular, through each point of M there passes a unique integral manifold of I. Proof. By Theorem 3.6 we have s1 = · · · = sn = 0, and since there are no integrability conditions the structure equations are dθa ≡ 0

mod {θa }.

There is another consequence of this corollary. Because of its many uses we state the result as a theorem. Theorem 3.12. Let (I, Ω) be a C ∞ exterior differential system and assume that: (i) the complex characteristic variety is empty, (ii) (technical assumption) the process of prolongation makes sense (i.e., at each stage we get a locally finite union of manifolds; this is automatic in the real analytic case). Then a prolongation of (I, Ω) is either empty or is a Frobenius system. In particular, for a suitable q each connected integral manifold of (I, Ω) is uniquely determined by its q-jet at one point. Informally, we may say that in case ΞC = ∅ the integral manifolds of (I, Ω) depend on a finite number of constants. Proof. We will prove in Chapter VI that a suitable prolongation (I (q) , Ω) of (I, Ω) is either empty or involutive. Since the integral manifolds of (I, Ω) and (I (q) , Ω) are locally in one-to-one correspondence, we may restrict to the latter case. From the remarks following (79) above we infer that the complex characteristic variety of (I (q) , Ω) is empty. By Corollary 3.11, (I (q) , Ω) is a Frobenius system. Its integral manifolds are uniquely specified by constants; through each point of M (q) there is a unique connected integral manifold. This translates into the assertion that connected integral manifolds of (I, Ω) are uniquely determined by their q-jets at one point. Example 3.13. We shall give an example of the finiteness Theorem 3.12 for a linear P.D.E. system that arose initially in algebraic geometry. This example is for illustrative purposes and will not be referred to elsewhere in the book. Let E → M be a vector bundle over a manifold, E the sheaf of C ∞ sections of E, and Θ ⊂ E a subsheaf. We ask for another vector bundle F → M and linear differential operator (85)

D:E →F

whose kernel is Θ. A first candidate for F may be obtained as follows: For each k let J k (E) → M be the bundle of k-jets of sections of E → M , and denote by J k (E) the sheaf of C ∞ sections of this jet bundle. There is the universal k th order differential operator jk : E → J k (E)

186


that sends a section s of E to its k-jet jk (s) (in local coordinates, jk (s)(x) is the Taylor series up through order k of s at the point x). We then consider the images jk (Θ) ⊂ J k (E) of k-jets of sections of Θ, and we assume that the values jk (s)(x) (s ∈ Θ and x ∈ M ) form a sub-bundle J k (Θ) of J k (E). For some k for which J k (Θ) = J k (E) we set F = J k (E)/J k (Θ) and consider the k th order operator D:E →F defined by (Ds)(x) = πjk (s)(x) where π : J (E) → F is the projection. Clearly, Θ ⊂ ker D; in this way we obtain candidates for linear differential operators (85) whose solution sheaf is Θ. (It is also clear that any solution to this problem must essentially be of this form.) Now let G = G(k, V ) be the Grassmann manifold of k-planes in a vector space V (which may be real or complex; it doesn’t matter). Over G we have the universal sub-bundle S → G, the trivial bundle V˜ = G×V , and the universal quotient bundle Q → G, all fitting in the standard exact sequence k

(86)

0 → S → V˜ → Q → 0.

The constant sections V of V˜ → G project to give a subsheaf V ⊂ Q, and we ask for a linear differential operator D:Q→R

(87)

whose kernel is V (here we abuse notation by identifying V with a subsheaf of Q). More generally, for any submanifold M ⊂ G we may restrict (86) to M and ask the same question. It is this latter situation that arose in algebraic geometry. What we shall do here is: i) for any M ⊂ G define a linear, 1st order operator (87) such that V ⊂ ker D; ii) determine geometric conditions on M such that the complex characteristic variety ΞC of D is empty. By Theorem 3.12 this implies that: Over any open subset U ⊂ M , ker D is a finite-dimensional subspace of H 0 (U, E); and iii) show that if M = G, then over any open subset U ⊂ G, ker D = V . We note that since the symbol of D will not be identically zero, the linear P.D.E. system Ds = 0 cannot be involutive (this is a consequence of Corollary 3.11 above). We begin with some notation. Given M ⊂ G and a point x ∈ M we denote by Sx ⊂ V the corresponding k-plane and set Qx = V /Sx . Since the question is local we may choose a frame e1 (x), . . . , eN (x) (N = dim V ) for V˜ → N such that e1 (x), . . . , ek (x) is a frame for S → M . Using the range of indices

1 ≤ α, β, γ ≤ k, k + 1 ≤ µ, ν ≤ N

we set (88)

deα = ωαβ eβ + ωαµ eµ deµ = ωµα eα + ωµν eν .


187

For each x ∈ M the map Tx M → Hom(Sx , Qx ) = Sx∗ ⊗ Qx given for v ∈ Tx M by

v → ωαµ (x), ve∗α (x) ⊗ eµ (x)

is well-defined and gives an inclusion (89)

T M ⊂ Hom(S, Q).

When M is an open set on G this inclusion is an equality and gives the well known identification T G ∼ = Hom(S, Q). We now fix a point x0 ∈ M and set T = Tx0 M , S = Sx0 , Q = Qx0 ; we thus have a subspace T ⊂ Hom(S, Q). For this subspace we consider the following two conditions: ' ker ϕ = (0); (90) ϕ∈T

(91)

For any hyperplane H ⊂ T ' ker ϕ = (0). ϕ∈H

If the condition (90) is satisfied at each point then we shall prove that J 1 (V ) ⊂ J 1 (Q) is a proper sub-bundle. We then set R = J 1 (Q)/J 1 (V ) and define (92)

D:Q→R

by the above procedure (D = πD1 where π : J 1 (Q) → R is the projection). This is a linear, 1st order differential operator and we shall prove that: (93)

If (91) is satisfied, then the complex characteristic variety ΞC of D is empty.

As noted above, this implies that locally ker D is a finite-dimensional vector space. (94)

If dim V ≥ k + 2 and M is an open set of G then (90) and (91) are satisfied, and in fact ker D = V .

The dimension restriction simply means that G is not a projective space. To get some feeling for the conditions (90) and (91) we set dim V = N and assume that k ≤ N − k; i.e., dim S ≤ dim Q. Then it is easy to see that: Condition (90) is generic if dim M ≥ 1. Condition (91) is generic if either k ≤ N − k − 1 and dim M ≥ 2, or k = N − k and dim M ≥ 3.

188


We are now ready to compute. If v ∈ V gives a constant section of Q, then writing v = vα eα + vµ eµ we have 0 = dv ≡ (dvµ + vλ ωλµ + vα ωαµ )eµ

mod {eα }

which implies that dvµ + vλ ωλµ + vα ωαµ = 0.

(95) Now the map

V → J 0 (Q)x = Qx

(96)

is obviously surjective for each x ∈ M . We shall show that if (90) is satisfied then the map V → J 1 (Q)x is injective for each x ∈ M (in fact, this is equivalent to (90)). For this we may assume that v ∈ Sx (i.e., v = 0 in Qx ), so that all vλ (x) = 0. Then (95) gives dvµ (x) = −vα (x)ωαµ (x).

(97) Now in the sequence

ρ

0 → Tx∗ M ⊗ Qx → J 1 (Q)x − → J 0 (Q)x → 0, ρ(v) = 0, and so the 1-jet j1 v(x) is given by dvµ (x)eµ (x) ∈ Tx∗ M ⊗ Qx . By (97) this is zero if, and only if, vα (x)ωαµ (x) = 0 for all µ. Clearly this is just a reformulation of (90). Next we want to compute the symbol σ of the operator (92). Working over a fixed point x0 ∈ M and using our above notations, the symbol is a map (98)

σ : T∗ ⊗ Q → R

where R is the fibre of J 1 (Q)/J 1 (V ) over x0 . To identify R, we note that by the surjectivity of (96) it is a quotient space of T ∗ ⊗ Q. Denoting by j : T ⊂ S∗ ⊗ Q the inclusion (89) over x0 , we define a linear mapping λ : S → T ∗ ⊗ Q by the commutative diagram S idS ⊗ I

λ

− → S ⊗ Q∗ ⊗ Q

T∗ ⊗ Q j ∗ ⊗ idQ

where I ∈ Q∗ ⊗ Q = Hom(Q, Q) is the identity. Our main observation is the following consequence of (95): R = T ∗ ⊗ Q/λ(S).


189

Thus the symbol (98) is just the quotient mapping T ∗ ⊗ Q → T ∗ ⊗ Q/λ(S). The characteristic variety is non-empty if, and only if, we have λ(s) = η ⊗ q

(99) ∗

for some 0 = s ∈ S, η ∈ T , q ∈ Q. Setting H = η ⊥ ∩ T, (99) is easily seen to be equivalent to ' s∈ ker(j(ϕ)). ϕ∈H

Comparing with (91) we obtain a proof of (93). We will now prove (94). Let q = vµ eµ be a section of Q over an open set in G and assume that Dq = 0. By definition this means that there exists s = vα eα such that, for each µ, dvµ + vλ ωλµ + vα ωαµ = 0.

(100)

The exterior derivatives of (88) give dωij = ωik ∧ ωkj ,

1 ≤ k, j, k ≤ N.

Using this the exterior derivative of (100) gives dvλ ∧ ωλµ + vλ ωλα ∧ ωαµ + vλ ωλν ∧ ωνµ + dvα ∧ ωαµ + vα ωαβ ∧ ωβµ + vα ωαλ ∧ ωαµ = 0. Plugging (100) into this equation several cancellations occur and it becomes (dvα + vλ ωλα + vβ ωβα ) ∧ ωαµ = 0 for µ = k + 1, . . . , N . Now all the forms ωαµ are linearly independent (they give a local coframe for G), and the Cartan lemma implies that dvα + vλ ωλα + vβ ωβα ∈ span{ω1µ , . . . , ωkµ } for each µ = k + 1, . . . , N . If N ≥ k + 2 then we may choose µ = ν and use span{ω1µ , . . . , ωkµ } ∩ span{ω1ν , . . . , ωkν } = (0) to conclude that (101)

dvα + vλ ωλα + vβ ωβα = 0.

For the V -valued function v = vα eα + vµ eµ (100) and (101) imply that dv = 0.

The proof shows that if N ≥ k + 2 and M ⊂ G is a generic submanifold with dim M ≥ 2k, then ker D = V . (Note that dim G = k(N − k) so that M must be an open set on G if N = k + 2.) Unfortunately, for the cases that arise in algebraic geometry we have that approximately dim M = k, so that prolongation is necessary to decide if Ker D = V .

190


Returning to the general discussion, in section 2 above we have defined the subbundle S ⊂ L∗ to be the image of the Cauchy characteristics and have proved (cf. equation (19) there) that ΞC ⊂ PSC⊥ . In the involutive case there is a converse. To explain it, for each subset Σ ⊂ PLC whose fibres Σx are algebraic subvarieties of PLx,C ∼ = Pn−1 , we define the span of Σx to be {Σx } =

' (linear spaces containing Σx )

and set {Σ} =

{Σx }.

x∈M

Theorem 3.13. In case (I, Ω) is involutive, we have SC⊥ = {ΞC}. In the extreme case when ΞC is empty, this gives SC⊥ = LC so that (I, Ω) is a Frobenius system, which is Corollary 3.11 above. We shall not prove this result in this book. (v) The characteristic variety and K-singular integral elements. As we have defined it, the characteristic variety essentially has to do with characteristic, or singular, hyperplanes in n-dimensional integral elements. On the other hand, if (I, Ω) is an involutive Pfaffian differential system of character l then the uniqueness of extensions in the Cauchy problem for n-dimensional integral manifolds occurs along l-dimensional submanifolds. Thus, we may expect that the characteristic variety should consist of singular l-dimensional integral elements. In other words, when l < n − 1 (i.e., roughly speaking when we are in the overdetermined case) there are two possible characteristic varieties, and it is obviously important to relate them. To explain this we recall that Gp(I) ⊂ Gp(T M ) denotes the set of p-dimensional integral elements of a differential ideal I on manifold M , and that the rank of the polar equations at (x, E) ∈ Gp (I) is denoted by ρ(E). Next, we recall that (i) (x, E) ∈ Gp (I) is K-regular if near (x, E) the set Gp (I) is a manifold with defining equations ϕ|E = 0, (x , E ) ∈ Gp (T M ) for ϕ ∈ I, and (ii) ρ(E ) is constant near (x, E). If (x, E) is not K-regular then it is said to be K-singular. Now suppose that (I, Ω) is a linear Pfaffian differential system given by a filtration I ⊂ J ⊂ T ∗ M and with characteristic variety Ξ ⊂ PL


191

where L = J/I with rank L = n. For any n-dimensional integral manifold N ⊂ M , the restriction J → T ∗ (N ) induces isomorphisms L|N ∼ = T ∗N PL|N ∼ = Gn−1 (T N ), so that the complex characteristic variety ΞC maps to a set of hyperplanes in the tangent spaces to N that we think of as giving characteristic hyperplanes for a determined Cauchy problem posed along hypersurfaces in N . On the other hand, if (I, Ω) has character l then, from the proof of the Cartan–K¨ ahler theorem, there is a sequence of uniquely determined Cauchy problems beginning with one posed along general l-dimensional submanifolds of N . The characteristics for the first of these problems should appear in Gl (T N ), which leads us to look for some sort of characteristic variety in Gl (L∗ ) ∼ = Gn−l (L). In fact, for each p with 1 ≤ p ≤ n − 1 we shall now define Λp ⊂ Gp (L∗ ) with the properties that Λn−1 = Ξ and that Λl = Λ is the characteristic variety of primary interest. Using the identification Gp (L∗ ) ∼ = Gn−p(L) given by sending a p-plane E in an n-dimensional vector space to its (n − p)dimensional annihilator E ⊥ in the dual space, for each (x, E) ∈ Gp(L∗ ) we choose our basis ω1 , . . . , ωn for L so that ωp+1 (x), . . . , ωn (x) gives a basis for E ⊥ . Consider now the tableau matrix ⎡ 1 ⎤ π1 (x) . . . πn1 (x) ⎢ ⎥ .. π(x) = ⎣ ... ⎦ mod Jx , . π1s0 (x) . . . πns0 (x) and denote by σ(E) the number of 1-forms in the first p columns of π(x) that are linearly independent mod Jx . Under a substitution p+1≤λ≤n

ω ˜ λ = ωλ ρ

ω ˜ =ω

ρ

+ Aρλ ωλ

1≤ρ≤p

we have π ˜ρa = πρa π ˜λa = πλa + Aρλ πρa , from which it follows that σ(E) is well-defined. Definition 3.14. i) We define Λp ⊂ Gp (L∗ ) by Λp = {E : σ(E) < s1 + · · · + sp } ii) If (I, Ω) has character l, then we define the Cartan characteristic variety Λ to be Λl .

192


We may intuitively think of Λp as the set of K-singular p-dimensional integral elements. In fact, as is easily verified from the discussions in Chapter IV, the precise statement is this: If there are no integrability conditions, so that the symbol relations may be assumed to be Baλi πia ≡ 0 mod I, then Λp is the set of p-dimensional integral elements whose polar equations have smaller rank than is generically the case. In particular, in the absence of integrability conditions, Λn−1 coincides with characteristic variety Ξ as given above in terms of the symbol. According to the proof of the Cartan–K¨ ahler theorem, the Cartan characteristic variety determines the set of integral elements that are characteristic for the last Cauchy–Kowaleski system in which there is any freedom in assigning initial data. It is clear how to define Λp,C and ΛC . A fundamental result is given by the following Theorem 3.15. In case (I, Ω) is involutive we have ΛC = {E ∈ Gl (L∗C ) : E ∈ [ξ]⊥ for some [ξ] ∈ ΞC } ΞC = {[ξ] ∈ PLC : E ∈ ΛC for all E ⊂ [ξ]⊥}. Simple examples show that the result is false without the assumption of involutiveness. Since we may have Ξ = ∅ but Λ = ∅ (see below), the result is false over R. What we can say is that the real Cartan characteristic variety Λ is given in terms of the complex usual characteristic variety ΞC by (102)

Λ = {E ∈ Gl (L∗ ) : E ⊂ [ξ]⊥ for some [ξ] ∈ ΞC }.

In particular, we may have Ξ = ∅ but ΞC and Λ both = ∅; see below. To picture Theorem 3.15 it may help to use the incidence correspondence Σ ⊂ Gl (L∗C ) × PLC defined by

Σ = {(E, [ξ]) : E ⊂ [ξ]⊥ }.

There are projections π1 Gl (L∗C )

Σ

π2 PLC

and the first assertion in Theorem 3.8 and (102) are equivalent to ΛC = π1 (π2−1 (ΞC )) Λ = ΛC ∩ Gl (L∗ ). Example 3.16. On R2m ∼ = Cm with coordinates z i = xi + 2m 2m structure J : R → R given by J(∂/∂xi ) = ∂/∂yi J(∂/∂yi ) = −∂/∂xi

√

−1 yi and complex


193

we consider the Cauchy–Riemann system (cf. Example 2.4 above). Since it is elliptic, the real characteristic variety Ξ = ∅. To describe the Cartan characteristic variety, since the system is translation invariant it will suffice to describe the fibre Λ0 of Λ over the origin, and using (102) this is given by Λ0 = {E ∈ Gm (R2m ) : E ∩ J(E) = (0)}. In other words, Λ0 consists of real m-planes E ⊂ R2m that contain at least one complex line (the latter being a real 2-plane F ⊂ E with J(F ) = F ). It is, of course, well known that real m-dimensional submanifolds Y m ⊂ Cm such that Ty (Y ) ∩ JTy (Y ) = (0) for every y ∈ Y are locally determining sets for holomorphic functions. In general we have Λp,0 = {E ∈ Gp(R2m ) : dim E ∩ J(E) ≥ max(1, 2(p − m) + 1)}. For instance, Λ2m−1,0 = G2m−1 (R2m ) contains no information. The second assertion in Theorem 3.15 gives ΞC in terms of ΛC as follows ΞC = {[ξ] : π2−1 ([ξ]) ⊂ π1−1 (ΛC )}. In other words, an (n − 1)-plane is characteristic only if every l-plane contained in it is Cartan characteristic. (vi) Integrability of the characteristic variety. (a) Let N be a manifold and Σ ⊂ PT ∗ N a subset. There is an associated eikonal equation EΣ defined as follows: A function ϕ(y) on N is a solution of EΣ if [dϕ(y)] ∈ Σy whenever dϕ(y) = 0. ˜ ⊂ T ∗ N be defined by More precisely, we let Σ ˜ = π −1 Σ ∪ {0} Σ where π : T ∗ N \{0} → PT ∗ N is the projection and {0} ⊂ T ∗ N is the zero section. ˜ is a conical subvariety of T ∗ N , i.e., it is invariant under the natural R∗ Then Σ action on T ∗ N . Moreover, any conical subvariety is of this form. If y1 , . . . , yn = (yi ) are local coordinates on N with induced coordinates (yi , ξi ) on T ∗ N , then we shall ˜ is defined by equations always assume that Σ is a subset with the property that Σ (103)

F λ (yi , ξi ) = 0

λ = 1, . . . , R

where the F λ are either C ∞ or real analytic functions depending on the category in which we are working. Definition 3.17. The eikonal equation is ∂ϕ(y) =0 F λ yi , (EΣ ) ∂yi

λ = 1, . . . , R.

194


Remarks. The functions F λ need only be defined microlocally, i.e., in open sets U ⊂ T ∗ N invariant under the natural R∗ action. In the cases of interest to us the fibres Σy of Σ → N will be algebraic varieties, so that the F λ (yi , ξi ) may be chosen to be homogeneous polynomials in the ξi whose coefficients are functions of the yi . In this case, the complexifications ΣC ⊂ PTC∗ N ˜ C ⊂ T ∗N Σ C

are naturally defined, and so the complex eikonal equation makes sense by allowing ˜ C . We remark the function ϕ to have complex values and requiring that dϕ ∈ Σ that we may have Σ = ∅ but ΣC = ∅. From now on we assume that the F λ (yi , ξi ) may be chosen to be homogeneous polynomials in the ξi . Definition 3.18. The subset ΣC ⊂ PTC∗ N is involutive in case the eikonal equation EΣC is involutive. To state the main result, we let I be a differential system on a manifold M and with complex characteristic variety ΞC ⊂ PUC∗ (cf. Definitions 1.1, 1.2 and the subsequent discussion). For any integral manifold f :N →M

(104)

of I there is the induced characteristic variety ΞC,N ⊂ PTC∗ N defined as follows: Given (104) we have a diagram PUC∗ ⏐ ⏐ "ω˜ (105)

fˆ∗

PTC∗ N −−−−→ Gn (T M ) ⏐ ⏐ ⏐ ⏐ " " N

−−−−→

M

f

ˆ f∗ (y) = f∗ (Ty N ) ω ˜ −1 (x, E) = PEC∗ . The condition that (104) be an integral manifold of I is that fˆ∗ (N ) ⊂ Gn (I), where

and the dotted arrow in (105) means that there is a natural mapping ω ˜ −1 (fˆ∗ (N )) → PT ∗ N. C

By definition, ΞC,N is the image of ΞC ∩ ω ˜ −1 (fˆ∗ (N )) under this mapping. Informally, we may say that ΞC,N is induced from the characteristic variety ΞC in each of the integral elements f∗ (Ty N ) ∈ Gn (T M ). Definition 3.19. We shall say that ΞC is involutive in case the eikonal equation EΞC,N is involutive for any integral manifold (104) of I. The main result is essentially the following: If I is involutive, then its characteristic variety is involutive. More precisely, the result is


195

Theorem 3.20. Let E ∈ Gn (I) be an ordinary integral element. Then ΞC is involutive in a (possibly smaller) neighborhood V of E. What this means is that ΞC,N is involutive for all integral manifolds (104) satisfying fˆ∗ (N ) ⊂ V . We will prove this result only in the case when the characteristic varieties ΞC,E = ΞC ∩PEC are points. By Theorem 3.6 this corresponds to the case where the Cartan characters are given by (∗)

s1 = s0 , s2 = · · · = sn = 0.

Referring to (97) in §5 of Chapter IV we see that the symbol relations given by (96) in that section are given by commuting matrices Cρ . We shall make the additional assumption that (∗∗)

the Cρ are simultaneously diagonalizable.

It is interesting to note that Cartan stated the above theorem under the assumption (∗) (cf. Cartan [1953]9). He also proved the result under our additional assumption (∗∗), and by the computation of several examples he showed that the result is much more subtle in case the Cρ may have non-diagonal Jordan normal forms. More recently in Guillemin, Quillen and Sternberg [1970] Theorem 3.20 is stated and proved for involutive P.D.E. systems. There they also make a technical assumption analogous to but weaker than our assumption (∗∗), but there is no restriction on the dimension of the characteristic variety. Subsequently the general result was proved in Gabber [1981], where additional references may be found. Since not every exterior differential system is derived from a P.D.E. system, the Guillemin, Quillen and Sternberg result does not immediately imply Theorem 3.20. However, the result is not really in doubt; even the stronger theorem corresponding to the result proved by Gabber is certainly true. What is important, in our opinion, is that we do not know a proof of Theorem 3.20 using moving frames in the spirit of the one we shall give below in our special case. That argument will show that the theorem falls out by exterior differentiation of the structure equations of a involutive system, where the assumption (∗∗) is used to put these equations in a particular form. We think it is a very worthwhile problem to give a similar proof of the full result. It will simplify our notations if we now work only with the real characteristic variety and observe that the arguments remain valid in the complex case. Unless mentioned to the contrary this will now be done. ˜ (b) We will now derive conditions for the involutivity of a conical submanifold Σ ∗ λ i of T N . For this we assume that we may micro-locally choose functions F (y , ξi ) ˜ We denote by {f, g} such that (103) gives a regular set of defining functions for Σ. the Poisson bracket of functions locally defined on T ∗ N , and recall that by definition (106)

{f, g} =

∂f ∂g ∂f ∂g − i . i ∂ξi ∂y ∂y ∂ξi i

9 The result is announced on page 1127 in Part II of the 1984 edition of the collected works. Numerous special cases were worked out in the paper preceding the announcement.

196


More intrinsically, if (107)

η = ξi dyi

is the tautological 1-form on T ∗ N with associated symplectic form (108)

Θ = dη = dξi ∧ dyi ,

then {f, g}Θn = n df ∧ dg ∧ Θn−1 . It is well known that the condition for the involutivity of EΣ is (109)

˜ {F λ , F µ} = 0 on Σ,

1 ≤ λ, µ ≤ R.

We shall derive this result from differential system point of view. This computation will also allow us to reformulate the condition for integrability in a way that leads naturally to the proof of Theorem 3.20. The result is local (even micro-local), and so we work in the above local coordinates and consider T ∗ N ×R as J 1 (Rn , R) with coordinates (y1 , . . . , yn , ψ, ξ1 , . . . , ξn ). ˜ × R defined by In J 1 (Rn , R) we consider the submanifold P ∼ =Σ F λ (yi , ξi ) = 0

λ = 1, . . . , R.

On P we have the contact system (J , Φ) generated by the equations (110)

θ = dψ − ξi dyi = 0 Φ = dy1 ∧ · · · ∧ dyn = 0.

Lemma 3.21. Locally on P we may choose a coframe ϕ1 , . . . , ϕn , θ, πR+1 , . . . , πn such that Φ ≡ ϕ1 ∧ · · · ∧ ϕn mod {θ} and (111)

1 dθ ≡ πε ∧ ϕε + cij ϕi ∧ ϕj mod {θ} 2

where ε = R + 1, . . . , n and cij + cji = 0. Proof. The structure equation of (110) is (112)

dθ ≡ −dξi ∧ dyi

mod {θ},

and the symbol relations on the dξi are (113)

∂F λ (yi , ξi ) dξi ≡ 0 mod {θ, dyi }. dξi

By our assumption, the matrix ∂F λ /∂ξi has everywhere rank R, and so we may find invertible matrices Aλµ and Bji such that Aλµ ∂F µ /∂ξj Bji = δiλ .


197

Setting dξi = −Bij πj the symbol relations (113) imply that πλ ≡ 0 mod {θ, dyi },

(114)

λ = 1, . . . , R.

Then for ϕi = Bji dyj we have −dξi ∧ dyi = Bij πj ∧ dyi = πj ∧ ϕj , so that by (112) and (114) dθ ≡ πε ∧ ϕε

mod {θ, ϕi }.

Writing this out modulo {θ} gives (114) for suitable functions cij = −cji .

To establish (109), we consider a differential system (J , Φ) with structure equations on an open set U ⊂ R2n−R+1

(115)

(i) θ = 0 (ii) dθ ≡ πε ∧ ϕε + 12 cij ϕi ∧ ϕj mod {θ},

cij + cji = 0

(iii) Φ = ϕ ∧ · · · ∧ ϕ = 0 1

n

where ϕ1 , . . . , ϕn , θ, πR+1 , . . . , πn are a local coframe. We may make a substitution (116)

πε → πε − pεi ϕi ,

pεδ + pδε = 0,

without effecting the form of the structure equations (115). When this is done we may eliminate the terms cεδ , cεµ

1 ≤ ε, δ ≤ R and R + 1 ≤ µ ≤ n

in dθ, and then (117)

1 dθ ≡ πε ∧ ϕε + cλµ ϕλ ∧ ϕµ 2

mod {θ}.

It is easy to see that the 2-form T =

1 cλµ ϕλ ∧ ϕµ 2

is, up to a conformal factor, invariantly associated to (J , Φ). Indeed, the only substitution (116) that leaves the form (117) invariant is when all pεi = 0, and consequently the condition that the integrability conditions be satisfied for (J , Φ) is (118) In this case (J , Φ) is involutive.

cλµ = 0.

198


In fact, assuming (118) we may “integrate (J , Φ) by O.D.E.’s” as follows: Denoting by {∂/∂ϕi , ∂/∂θ, ∂/∂πε } the dual frame to {ϕi , θ, πε}, the vector fields ∂/∂ϕλ satisfy ∂/∂ϕλ

θ=0 dθ ≡ cλµ ϕµ mod {θ}.

∂/∂ϕλ

Thus, assuming (118) the vector fields ∂/∂ϕλ are Cauchy characteristics, and in fact it is easy to see that, in the notation of Chapter I, A(Φ) = span{∂/∂ϕ1 , . . . , ∂/∂ϕR }. Local integral manifolds for (J , Φ) may be found by prescribing arbitrary values of ψ(x) along an Rn−R ∩ U in a general linear coordinate system, and then flowing this initial data out along the R-dimensional integrable distribution C(Φ), just as is done in the case R = 1 as explained in Chapter II. It remains to mutually identify the conditions (109) and (118). Both are in˜ ⊂ T ∗ N and do not depend on the local variantly attached to the submanifold Σ defining equations or choice of coframing. It will therefore suffice to identify these ˜ two sets of conditions at a point (y i , ξi ) ∈ Σ. For this we will show that (109) expresses the condition that there be integral elements of (J , Φ) at each point of P . We may choose the F λ (yi , ξi ) such that ∂F λ i (y , ξi ) = δjλ . ∂ξj

(119)

We view integral elements of (J , Φ) as n-planes E n in T(yi ,ξ ) (T ∗ N × R), and by i (110) and (119) any E n on which Φ = 0 has equations ⎧ ∂F λ ⎪ ⎪ + (y i , ξi )dyj = 0 λ = 1, . . . , R dξ ⎪ λ ⎪ ⎪ ∂yj ⎪ j ⎪ ⎪ ⎨ dξρ + pρj dyj = 0 ρ = R + 1, . . . , n (120) ⎪ ⎪ j ⎪ ⎪ ⎪ ⎪ ⎪ ξ i dyi ⎪ ⎩ dψ = i

where pρj are to be chosen to annihilate dθ. By (112) the restriction of dθ to the n-plane (120) is (121)

∂F λ λ,j

∂yj

dyj ∧ dyλ +

pρj dyj ∧ dyρ ,

ρ,j

where the ∂F λ /∂yj are to be evaluated at (y i , ξ i ). We may set pρσ = 0 and pρλ = −∂F λ /∂yρ so that (121) reduces to λ,µ

∂F λ /∂yµ dyµ ∧ dyλ .


199

Thus the vanishing of all (∂F λ /∂yµ )(y i , ξi ) = {F µ , F λ }(yi , ξ i )

(122)

is equivalent to the existence of an integral element of (J , Φ) at (y i , ξ i ) and this establishes the equivalence of (109) and (118). (c) In preparation for the proof of Theorem 3.7 we want to express the involu˜ ⊂ T ∗N tivity conditions (109) in more geometric form. For this we recall that Σ has dimension given by ˜ = 2n − R, dim Σ and we denote by Θ the symplectic form (108) on T ∗ N . Proposition 3.22. The involutivity condition (109) is equivalent to rank Θ = 2(n − R)

(123) ˜ on Σ.

˜ Proof. It will suffice to verify the equivalence of (109) and (123) at a point of Σ where (119) is satisfied. Then, as in (121), Θ=−

∂F λ λ,j

=

∂yj

(dξρ +

ρ

=

dyj ∧ dyλ +

dξρ ∧ dyρ

ρ

∂F λ λ

γρ ∧ dyρ +

ρ

∂yρ

dyλ ) ∧ dyρ +

λ,µ

∂F λ λ,µ

∂F λ

∂yµ

∂yµ

dyλ ∧ dyµ

dyλ ∧ dyµ

λ λ ρ λ ∗˜ where the forms γρ = dξρ + λ ∂F ∂y ρ dy , dy , dy give a coframe for T Σ at the point in question. It follows from this last expression and (122) that (123) is equivalent to (109). A noteworthy special case arises when R = n−1, which is the case corresponding to our assumption (∗). Then Σ ⊂ PT ∗ N consists of points, say d of these, in each fibre Σy lying over y ∈ N . Geometrically, each of these d points gives a hyperplane Hα (y) in Ty N , and we claim that (124)

˜ is equivalent to the integrability The involutivity of Σ of each of the distributions Hα .

Proof. It will suffice to locally treat the case d = 1. Then Σy is given by [η(y)] where η(y) = ξi (y)dyi i ∗

is a non-zero section of T N , and ˜ = {λη(y) : λ ∈ R and y ∈ N }. Σ

200


˜ On Σ Θ = d(λη(y)) = dλ ∧ η(y) + λdη(y). If the integrability condition dη(y) = γ ∧ η is satisfied, then Θ = (dλ + λγ) ∧ η has rank 2. Conversely, if we locally choose forms β 1 , . . . , β n−1 such that span{η, β 1 , . . . , β n−1 } = span{dy1 , . . . , dyn }, then writing dη = γ ∧ η + β where β involves only the β i we have Θ = (dλ + λγ) ∧ η + β. From this it is apparent that Θ has rank 2 only if β = 0.

Proof of Theorem 3.20. (under the further hypotheses (∗) and (∗∗)) We may prolong I to obtain an involutive linear Pffafian system with the same induced characteristic variety on integral manifolds. Thus it will suffice to prove the result in this case. Let I ⊂ Ω1 (M ) be an involutive linear Pfaffian system and denote by {I} ⊂ ∗ Ω (M ) the algebraic ideal generated by I. Its structure equations may be written in the form dθα ≡ 0 mod {I} dθa ≡ πia ∧ ωi mod {I} where the θα span the 1st derived system, the forms θa with 1 ≤ a ≤ s span the remainder of I, and Ω = ω1 ∧ · · · ∧ ωn is the independence condition. Our assumption (∗) means that, if we set π1a = π a and denote by π = (π a ) the first column of the tableau matrix and by πρ = (πρa ) the ρth column for 2 ≤ ρ ≤ n, then the 1-forms π a are linearly independent modulo {θα , θa , ωi } and the symbol relations are πρ ≡ Bρ π mod J for s × s matrices Bρ . The integrability conditions mean that we may assume this equation holds modulo I. The assumption of involutivity is then equivalent to commutation relations [Bρ , Bσ ] = 0. Our assumption (∗∗) implies that we may make a linear change among the π a ’s so that the Bρ are all diagonal. We note that the πia may now be complex valued. Thus we have πρa ≡ λaρ π a mod I


201

(no summation), and if we set λa1 = 1 and ω ˆ a = λai ωi , then the characteristic variety is given by the s points [ˆ ωa ] ∈ PLC . The structure equations of I are dθα ≡ 0

(125)

mod {I}

ˆ dθ ≡ π ∧ ω a

a

a

mod {I}

(no summation), and by (124) we must show that (126)

On any integral manifold N we have a a dˆ ωN ≡ 0 mod ω ˆN .

Here, we recall our notation ψN = ψ|N for any submanifold N ⊂ M . The idea is to differentiate the equations (125), and in so doing we shall make use of a trick that was employed by Cartan. Namely, fixing N we may make a change πia → πia − paij ωj so that (πia )N = 0.

(127)

This does not mean that we adjoin the equations (127) to I; it means only that on the particular integral manifold N we may assume that (127) holds. We now write out the second equation in (125) as ˆ a + ϕaβ ∧ θβ + ϕab ∧ θb . dθa = π a ∧ ω Exterior differentiation and use of (125) gives for each fixed a ωa + ϕab ∧ π b ∧ ω ˆ b ≡ 0 mod {I, ω ˆ a }. π a ∧ dˆ

(128)

In any case we have expansions of the form a c ϕab ≡ Fbc π + Fbia ωi mod I

(129)

1 a i 1 a b π ∧ ωi + Aabc π b ∧ π c mod {I}. Cij ω ∧ ωj + Bbi 2 2 a a = −Cji satisfy suitable conditions. Using the trick (127) we must show that the Cij The ω-quadratic terms in (128) give (no summation on a) 1 a a i ω ∧ ωj + Fbia ωi ∧ π b ∧ ω ˆ b ≡ 0 mod {I, ω ˆ a }. π ∧ Cij 2 For each fixed b = a this implies that dˆ ωa ≡

ˆ b ≡ 0 mod {I, ω ˆ a }, Fbia ωi ∧ π b ∧ ω and the remaining equation becomes (no summation on a) 1 a a i ω ∧ ωj ≡ 0 mod {I, ω ˆ a }. π ∧ Cij 2 This gives 1 a i ˆ a} C ω ∧ ωj ≡ 0 mod {I, ω 2 ij which when plugged into (129) becomes ˆ a , π 1 , . . . , π s }. dˆ ωa ≡ 0 mod {I, ω Using (127) this implies our desired result (126).

202

VI. Prolongation Theory

CHAPTER VI PROLONGATION THEORY

As has been seen in earlier chapters, it often happens that a given differential system with independence condition fails to be involutive. The process of prolongation is designed to remedy this situation and will be discussed in this chapter. At the P.D.E. level, the process of prolongation is nothing more than introducing the partial derivatives of the unknown functions as new variables and then adjoining new P.D.E. to the original P.D.E. system which ensure that the new variables are, in fact, the partial derivatives of the original unknown functions. The objective in doing this is that it may happen that the new system of P.D.E. is involutive even though the original system is not. (For an explicit example of this, see Examples 1.1 and 1.2 in Section 1 below.) Geometrically, for an exterior differential system, prolongation is essentially the process of replacing the original exterior differential system I ⊂ Ω∗ (M ) by the canonical Pfaffian system with independence condition (I (1), Ω) defined on the space Vn (I) of n-dimensional integral elements of I. This is made precise in Section 1 under the assumption that the space Vn (I) is sufficiently “well-behaved”. More precisely, we assume that Vn (I) has a stratification into smooth submanifolds of Gn (T M ), an assumption which is always satisfied in practice or when I is real analytic. The remainder of Section 1 is devoted to three examples which illustrate several phenomena which may arise during the process of prolongation. In Section 2, we investigate the effect that prolongation has on a component Z of Vn (I) which consists of ordinary integral elements. We prove the expected result that the prolongation (I (1), Ω) is involutive on Z. Moreover, (see Theorem 2.1) we show that the Cartan characters of (I (1), Ω) on Z can be computed by the expected formula from the Cartan characters of Z as a component of Vn (I). This result is to be found in Cartan’s work in the case that I is a Pfaffian system in linear form. The more general case (which does not follow from the Pfaffian system case) is due to Matsushima [1953]. For the (simpler) proof in the Pfaffian system case, the reader may want to compare the discussion at the end of Chapter VIII, Section 2, where Cartan’s original argument (albeit in more modern language) is given. There remains the question of the effect of prolongation on a non-involutive exterior differential system. It was a conjecture of Cartan (based on his having computed a large number of examples) that, for any real analytic differential system I, a finite number of iterations of the process of prolongation applied to I would lead either to an involutive differential system or else to a system with no integral elements (and hence, no integral manifolds). Although Cartan made attempts to prove this important result (for example, see Cartan [1946]), he was never able to do so. It was Kuranishi [1957] who first provided a proof of Cartan’s conjecture under certain technical hypotheses on I, which, in practice, were always fulfilled. Since then, various improvements in the technical hypotheses have been made, (cf. Goldschmidt [1968a, 1968b] and Theorem 1.14, Chapter X), although the general result remains open. Since the technical hypotheses are difficult to check without computing the successive prolongations up to a certain order, the general Cartan–

§1. The Notion of Prolongation

203

Kuranishi prolongation theorem has so far been of more theoretical importance than practical importance. On the other hand, the fully linear theory is better behaved and the reader should consult Chapter X for that case. In any case, the validity of the theorem (even with the technical hypotheses) serves, in practice, as motivation for the computation of the prolongations. In the last two sections of this chapter, we discuss (a version of) the Cartan– Kuranishi prolongation theorem. In Section 3, we introduce the important notion of a prolongation sequence, which slightly generalizes the case of a sequence of prolongations for which the space of integral elements at each stage forms a smooth submanifold of the appropriate Grassmann bundle for which the projection to the appropriate base manifold is a submersion. In the terminology of Chapter IV, this corresponds to the case of a sequence of prolongations for which the “torsion” is always “absorbable”. We prove that, in this case, after a finite number of steps, the remaining differential systems in the sequence are all involutive (see Theorem 3.2). The crucial step is the reduction of the problem to a commutative algebra statement which is related to the Hilbert syzygy theorem (an important step in all of the known proofs of the Cartan–Kuranishi theorem). This commutative algebra statement is then proved in Chapter VIII. Finally, in Section 4, we relate Theorem 3.2 to the “classical” version of the Cartan–Kuranishi Prolongation Theorem, namely the case where the torsion is always absorbable. We then enter into a discussion of the general case and point out some of the difficulties and what is expected to be the general result. The upshot of our discussion is that, when dealing with a non-involutive differential system, the process of prolongation is an essential step in the study of the integral manifolds. Moreover, in practice, in the analytic category, the process satisfactorily answers the existence question for integrals of an exterior differential system. §1. The Notion of Prolongation. We begin by recalling some relevant constructions from earlier sections. Let M be a smooth manifold of dimension m and let n ≤ m be an integer. We let π : Gn (T M ) → M denote the Grassmann bundle whose fiber at x ∈ M consists of the space of n-planes E ⊂ Tx M . Every smooth immersion f : N n → M induces a canonical smooth map f1 : N n → Gn (T M ) by the formula f1 (p) = f∗ (Tp N ) ⊂ Tf(p) M . This f1 is clearly a lifting of f. The dimension of Gn (T M ) is m+n(m−n) and it carries a canonical Pfaffian system I of rank m − n which has the property that for every immersion f : N n → M the induced lifting f1 : N n → Gn (T M ) is a integral manifold of the exterior differential system I generated by I. Moreover, a smooth map ϕ : N n → Gn (T M ) is of the form ϕ = f1 for some immersion f : N n → M if and only if π ◦ ϕ : N n → M is an immersion and ϕ is an integral of the system I. In this case, it then follows that ϕ = (π ◦ ϕ)1 . Alternatively, we note that there is a canonical rank n independence condition Ω on Gn (T M ) with the property that the integrals ϕ : N n → Gn (T M ) of (I, Ω) are precisely the canonical lifts of immersions f : N n → M . For a more explicit description of this system, we refer the reader to Chapter IV. Let S ⊂ Gn (T M ) be a subset. We say that an immersion f : N n → M is an S-immersion if f1 (N ) ⊂ S. In most cases of interest, S will be a submanifold, so let us assume this for the moment. It is clear that a map ϕ : N n → S will be the

204


canonical lift of an S-immersion if and only if it is also an integral of the system (I, Ω). Thus, the study of S-immersions is equivalent to the study of the integrals of the system (I|S , Ω|S ) on S. A special case will be of great importance. If I ⊂ Ω∗ (M ) is a closed differential ideal and S = Vn (I), then the system (I|S , Ω|S ) is what we would like to call the (first) prolongation of I. The difficulty with this as a definition is that the space Vn (I) often fails to be a smooth manifold. In practice, however, we can usually write Sβ (1) Vn (I) = β∈B

where {Sβ | β ∈ B} is a stratification of Vn (I0 ) into irreducible smooth components. (This is always possible when I is real analytic.) Then we may consider the exterior differential system (2) (I (1), Ω(1)) = (I|Sβ , Ω|Sβ ). β∈B

defined on the disjoint union of the strata Sβ as the (first) prolongation of I. Note that, on each component Sβ , the prolongation of I is always a Pfaffian system with independence condition. For a Pfaffian system with independence condition (I, Ω) the first prolongation is defined to consist of the canonical system with independence condition on Vn (I, Ω) restricted to the components of Vn (I, Ω). In practice, we are usually interested in the integrals of I (or (I, Ω)) which satisfy some additional conditions. This often has the effect of restricting our attention to a particular smooth stratum in Vn (I, Ω) anyway. Before going on to the general theory of prolongation, we will discuss several examples. These examples will be used to motivate the development of the theory in the later sections of this chapter. Example 1.1. Consider the differential system which describes the simultaneous solutions of the pair of differential equations (3)

∂2u ∂2u ∂2u ∂2u − 2 = − 2 = 0. 2 ∂x ∂y ∂x2 ∂z

These equations describe a submanifold M of J 2 (R3 , R). With the usual coordinates (xi , u, pi , pij = pji ) on J 2 (R3 , R), M is given by the equations p22 = p33 = p11 . Comparing this system with the example in Chapter IV about pairs of second order equations, we note that the symbol quadrics are spanned by {(ξ1 )2 − (ξ2 )2 , (ξ1 )2 − (ξ3 )2 } and, since these have no common divisor, the system (3) of P.D.E. is not involutive. The differential system I is the restriction of the contact system on J 2 (R3 , R) to M and thus is generated by the Pfaffian system I spanned by the four 1-forms ϑ = du − p1 dx1 − p2 dx2 − p3 dx3 (4)

ϑ1 = dp1 − p11 dx1 − p12 dx2 − p13 dx3 ϑ2 = dp2 − p12 dx1 − p11 dx2 − p23 dx3 ϑ3 = dp3 − p13 dx1 − p23 dx2 − p11 dx3 .


205

The independence condition is given by Ω = dx1 ∧ dx2 ∧ dx3 . Setting πij = dpij and ωi = dxi , the structure equations become ⎤ ⎡ 0 ϑ ⎢ π11 ⎢ ϑ1 ⎥ d⎣ ⎦ ≡ −⎣ ϑ2 π12 ϑ3 π13 ⎡

(5)

0 π12 π11 π23

⎤ ⎡ ⎤ 0 ω1 π13 ⎥ ⎣ 2 ⎦ ⎦∧ ω π23 ω3 π11

mod I.

It is straightforward to compute that s0 = 4, s1 = 3, s2 = 1, and s3 = 0. By Cartan’s Test, the system will be involutive if the integral elements of (I, Ω) at each point of M form an (affine) space of dimension 5 = s1 + 2s2 + 3s3 . However, calculation yields that the integral elements at each point of M are described by 4 parameters. Explicitly, for any four real numbers r1 , r2 , r3 , r4 , the 3-plane based at any point of M which is annihilated by the 1-forms ϑ, ϑ1, ϑ2 , ϑ3 and the 1-forms ⎡

⎤ ⎡ ⎤ ⎡ ϑ4 π11 r1 ⎢ ϑ5 ⎥ ⎢ π12 ⎥ ⎢ r2 ⎣ ⎦=⎣ ⎦−⎣ ϑ6 π13 r3 ϑ7 π23 r4

(6)

r2 r1 r4 r3

⎤ r3 ⎡ 1 ⎤ ω r4 ⎥ ⎣ 2 ⎦ ⎦ ω r1 ω3 r2

is an integral element of (I, Ω) and every such integral element is of this form. Thus, V3 (I, Ω) = M × R4 and I (1) is generated on M (1) = M × R4 by the eight 1-forms {ϑ, ϑ1, ϑ2 , . . . , ϑ7 }. The structure equations of I (1) are dϑ ≡ dϑ1 ≡ dϑ2 ≡ dϑ3 ≡ 0 mod I (1) ⎤ ⎡ ⎤ ⎡ ⎤ π1 π2 π3 ϑ4 ω1 ⎢ π2 π1 π4 ⎥ ⎣ 2 ⎦ ⎢ ϑ5 ⎥ d⎣ ⎦ = −⎣ mod I (1) . ⎦∧ ω ϑ6 π3 π4 π1 3 ω ϑ7 π4 π3 π2 ⎡

(7)

where we have set πi = dri . The sequence of reduced Cartan characters of I (1) is easily computed to be (s0 , s1 , s2 , s3 ) = (8, 4, 0, 0). Moreover, the space of integral elements of (I (1) , Ω) at each point of M (1) can be seen to be parametrized by 4 (= s1 +2s2 +3s3 ) parameters t1 , t2 , t3 , t4 in such a way that the annihilator of the corresponding integral element at any point of M (1) is spanned by the eight 1-forms {ϑ, ϑ1, ϑ2 , . . . , ϑ7 } together with the four 1-forms ⎤ ⎡ ⎤ ⎡ π1 t1 ϑ8 ⎢ ϑ 9 ⎥ ⎢ π 2 ⎥ ⎢ t2 ⎦ =⎣ ⎦−⎣ ⎣ ϑ10 π3 t3 ϑ11 π4 t4 ⎡ (8)

t2 t1 t4 t3

⎤ t3 ⎡ 1 ⎤ ω t4 ⎥ ⎣ 2 ⎦ ⎦ ω . t1 ω3 t2

It then follows, by Cartan’s Test, that the system (I (1) , Ω) is involutive on M (1) with Cartan character sequence (s0 , s1 , s2 , s3 ) = (8, 4, 0, 0) even though (I, Ω) is not involutive on M . Since the integrals of (I (1), Ω) on M (1) and (I, Ω) on M are in one-to-one correspondence, we see that we may actually study the integrals of (I, Ω) by applying the Cartan–K¨ ahler theorem to (I (1), Ω). The fact that s1 = 4 is the last non-zero Cartan character of I (1) indicates that the “general” integral manifold of (I (1), Ω) should depend on four functions

206


of one variable. Note also that the characteristic variety in each integral element E of (I (1) , Ω) consists of the four points [dx1 ± dx2 ± dx3 ] in PE ∗ . Now, in this particular example, it is possible to explicitly write down the general solution of (3). This general solution takes the form (9)

u(x1 , x2 , x3 ) = f++ (x1 + x2 + x3 ) + f+− (x1 + x2 − x3 ) + f−+ (x1 − x2 + x3 ) + f−− (x1 − x2 − x3 )

where the functions f±± are four arbitrary functions of 1 variable. However, it is not usually possible to write down the “general” solution of a P.D.E. so explicitly. For example, the reader might try analyzing the system (10)

∂2u ∂2u ∂2u + λ1 u = + λ2 u = + λ3 u 2 2 ∂x ∂y ∂z 2

where the λi are arbitrary constants. Again, it turns out that the first prolongation on M (1) is involutive even though the natural system on M ⊂ J 2 (R3 , R) is not. The next example illustrates the fact that the prolongation process may need to be iterated several times before the resulting system becomes involutive. Example 1.2. Consider the pair of equations for u as a function of x and y ∂nu ∂nu = = 0. ∂xn ∂yn

(11)

In a departure from our usual notation for coordinates on J n (R2 , R), let us use pi,j to represent ∂ i+j u/∂xi ∂yj . Thus, p0,0 = u and the equations (11) correspond to the submanifold M ⊂ J n (R2 , R) given by the equations p0,n = pn,0 = 0. The contact Pfaffian system on J n (R2 , R) is then generated by the 1-forms {ϑi,j | i+j < n} where ϑi,j = dpi,j − pi+1,j dx − pi,j+1 dy.

(12)

We let I denote the restriction of this Pfaffian system to M . The structure equations of I then can be written as dϑi,j ≡ 0 for 0 ≤ i + j < n − 1

(13a) ⎛ dϑ (13b)

n−1,0

⎞

⎛

0

⎜ dϑn−2,1 ⎟ ⎜ πn−1,1 ⎜ ⎟ ≡ −⎜ . .. ⎝ ⎠ ⎝ . . . π1,n−1 dϑ0,n−1

πn−1,1 ⎞ πn−2,2 ⎟ dx .. ⎟ ⎠ ∧ dy . 0

where the congruences are taken mod I and πk,l = dpk,l for k + l = n. It is straightforward to compute that s1 = n − 1 and s2 = 0. Thus, in order for (I, Ω) to be involutive (where Ω = dx ∧ dy), it would be necessary for there to exist an (n − 1)-parameter family of integral elements of (I, Ω) at each point of


207

M . However, it is easy to see that there is only an (n − 2)-parameter family of integral elements of (I, Ω) at each point of M . In fact, M (1) = M × Rn−2 and we can introduce coordinates {pi,j | 2 ≤ i, j; i + j = n + 1} on the Rn−2 -factor so that the annihilator of the corresponding integral element is spanned by the 1-forms {ϑi,j | 1 ≤ i + j ≤ n − 1} and the 1-forms ⎛ϑ (14)

n−1,1

⎞

⎛π

n−1,1

⎞

⎛

0

⎜ ϑn−2,2 ⎟ ⎜ πn−2,2 ⎟ ⎜ pn−1,2 ⎜ . ⎟ ≡ ⎜ . ⎟−⎜ . ⎝ . ⎠ ⎝ . ⎠ ⎝ . . . . p2,n−1 ϑ1,n−1 π1,n−1

pn−1,2 ⎞ pn−2,3 ⎟ dx .. ⎟ ⎠ dy . . 0

Note that M (1) can be regarded as the submanifold of J n+1 (R2 , R) defined by the equations (15)

p0,n = pn,0 = p0,n+1 = p1,n = pn,1 = pn+1,0 = 0.

Under this identification, I (1) becomes the restriction to M (1) of the contact Pfaffian system on J n+1 (R2 , R). The structure equations of I (1) are easily seen to be dϑi,j ≡ 0

(16a) ⎛ dϑ (16b)

n−1,1

⎞

for 0 ≤ i + j < n ⎛

0

⎜ πn−1,2 ⎜ dϑn−2,2 ⎟ ⎟ ≡ −⎜ . ⎜ . ⎠ ⎝ . ⎝ .. . π2,n−1 dϑ1,n−1

πn−1,2 ⎞ πn−2,3 ⎟ dx .. ⎟ ⎠ ∧ dy . 0

where the congruences are taken mod I (1) and πk,l = dpk,l for k + l = n + 1. We compute that s1 = n − 2 and s2 = 0, but that, for n ≥ 3, there is only an (n − 3)-parameter family of integral elements of (I (1) , Ω) at each point of M (1) . Thus, for n ≥ 3, (I (1), Ω) is not involutive. It is easy to continue this process. If we inductively define (I (k+1), Ω) = ((I (k))(1) , Ω), we find that, for k ≤ n − 1, the system I (k) is diffeomorphic to the restriction of the contact Pfaffian system on J n+k (R2 , R) to a submanifold M (k) defined by the equations (17)

pi,j = 0 whenever n ≤ i + j ≤ n + k and max{i, j} ≥ k.

For this system, s1 = n−k −1 and s2 = 0. However, when k ≤ n−2, the dimension of the space of integral elements of (I (k), Ω) at any point of M (k) is only n − k − 2. Thus, the system (I (k), Ω) is not involutive on M (k) for k ≤ n − 2. However, it is also easy to see that the Pfaffian system I (n−1) on M (n−1) is a Frobenius system and hence the system (I (n−1), Ω) is involutive on M (n−1). Since the rank of I (n−1) is n2 , it follows that there is an n2 -parameter family of local solutions of the system (11). This was expected since the solutions of (11) are clearly the polynomials in x and y whose x-degree and y-degree are less than or equal to n − 1.

208


In the next example, we treat a problem involving the geometry of Riemannian submersions in which the corresponding Vn (I) has several components. Example 1.3 (Riemannian submersions). Let (N 3 , dx2 ) be a Riemannian manifold with constant sectional curvature K. Our problem is to classify the Riemannian submersions f : (N 3 , dx2 ) → (Σ2 , dσ 2 ) where (Σ2 , dσ 2 ) is a Riemannian surface. We do not specify the metric on Σ2 in advance. Given such an f, we may locally choose an orthonormal frame field e = (e1 , e2 , e3 ) on N so that e3 is tangent to the fibers of f. Then the hypothesis that f be a Riemannian submersion is equivalent to the condition that f ∗ (dσ 2 ) = (η1 )2 + (η2 )2 where η = (η1 , η2 , η3) is the coframing dual to the frame field e. It follows that there exists a “connection form” γ in the domain of e so that dη1 = γ ∧ η2 and dη2 = −γ ∧ η1 . Conversely, if η = (η1 , η2 , η3) is any orthonormal coframing on an open set U ⊂ N such that there exists a 1-form γ satisfying dη1 = γ ∧ η2 and dη2 = −γ ∧ η1 , then it is easy to see that (η1 )2 + (η2 )2 is a well defined quadratic form on Σ2 = U/F where F is the foliation of U by the integral curves of the vector field e3 on U , i.e., the integral curves of the system η1 = η2 = 0 on U . It follows that the projection f : U → U/F is a Riemannian submersion. For any local orthonormal coframing η = (η1 , η2, η3 ) on U ⊂ N , there exist unique 1-forms (known as the Levi–Civita connection forms) ηij = −ηji which satisfy the “symmetry” condition dηi = −ηij ∧ ηj . The following equations for γ (18)

−η12 ∧ η2 − η13 ∧ η3 = dη1 = γ ∧ η2 −η21 ∧ η1 − η23 ∧ η3 = dη2 = −γ ∧ η1

are satisfiable only if there exist functions a, b, c on U so that η13 = aη2 + bη3 (19)

η23 = −aη1 + cη3 .

Conversely, the existence of such functions is sufficient to yield a solution, namely γ = η21 + aη3 , to (18). Thus the search for the desired Riemannian submersions f is locally equivalent to the search for orthonormal coframings η which satisfy certain first order differential equations. Actually, the coframing η carries slightly more information since two coframings determine the same Riemannian submersion iff their corresponding e3 foliations are the same. We shall see what becomes of this ambiguity in what follows. Using the above discussion as our guide, we may now set up a differential system to find the desired framings as follows. Let F → N 3 denote the orthonormal frame bundle of (N 3 , dx2) and let {ωi , ωij = −ωji } denote the canonical and connection 1-forms on F . They satisfy the first and second structure equations of Cartan: (20)

dωi = −ωij ∧ ωj dωij = −ωik ∧ ωkj + Kωi ∧ ωj .

(Here, we use the summation convention. Recall that ds2 has constant sectional curvature K.)


209

A (local) framing e is a (local) section e : U → F for an open set U ⊂ N . The forms {ωi , ωij = −ωji } have the property that e∗ (ωi ) = ηi and e∗ (ωij ) = ηij where η = (η1 , η2 , η3 ) is the local coframing which is dual to e. As we have already said, we are seeking (local) framings e for which there exists functions a, b, c so that (19) holds. This motivates us to define M = F × R3 (where we use a, b, and c as linear coordinates on the R3 -factor) and let I be the Pfaffian system generated by the 1-forms ϑ1 = ω13 − aω2 − bω3

(21)

ϑ2 = ω23 + aω1 − cω3 .

It is clear that every framing e which satisfies (19) gives rise to a unique integral manifold Ne of I in M on which the form Ω = ω1 ∧ ω2 ∧ ω3 does not vanish. Conversely, if I is the differential system generated by I on M then every integral of (I, Ω) in M is locally of the form Ne for some framing e which satisfies (19). It is easy to compute that the structure equations of I are given by dϑ1 ≡ −π3 ∧ ω2 − π4 ∧ ω3

(22)

dϑ2 ≡ π3 ∧ ω1 − π5 ∧ ω3

mod I

where

(23)

π3 = da −2a(bω1 + cω2 ) 2 π4 = db + cω12 + (a − K)ω1 −b(bω1 + cω2 ) π5 = dc − bω12 + (a2 − K)ω2 −c(bω1 + cω2 ).

Notice that the structure equations (22) imply that I has a one-dimensional Cauchy characteristic system spanned by the vector field X on M which satisfies ω12 (X) = 1 and α(X) = 0 for α = ω1 , ω2 , ω3 , ϑ1, ϑ2 , π3 , π4, or π5 . It is easy to see that the flow of this Cauchy vector field generates an S 1 -action on M which corresponds to the rotation of a frame e which fixes the e3 -component. We could get rid of this S 1 -action by passing to the quotient (M × R3 )/S 1 and working with the corresponding Pfaffian system there, but this causes computational difficulties since the quotient manifold has no natural coframing. Instead, our approach will be to augment the independence condition to Ω+ = ω1 ∧ ω2 ∧ ω3 ∧ ω12 and look for integrals of the system (I, Ω+ ). This has the added advantage that the integrals of this system correspond essentially uniquely to the local Riemannian submersions, as the reader can easily see. Now by the structure equations (22), it is clear that the reduced Cartan character sequence of (I, Ω+ ) is (s0 , s1 , s2 , s3 , s4 ) = (2, 2, 1, 0, 0). In order to have involutivity, Cartan’s Test thus requires that there be a 4 (= s1 + 2s2 + 3s3 + 4s4 ) parameter family of integral elements of (I, Ω+ ) at every point of M . However, it is easy to see that the integral elements of (I, Ω+) at a point of M are parametrized by only 3 parameters. Namely, for any real numbers p, q, r, the 4-plane at each point of M which is annihilated by the 1-forms ϑ1 , ϑ2 and the 1-forms ⎡ (24)

⎤ ⎡ ⎤ ⎡ ϑ3 π3 0 ⎣ ϑ 4 ⎦ = ⎣ π4 ⎦ − ⎣ 0 ϑ5 π5 −p

⎤⎡ ⎤ 0 p ω1 p q ⎦ ⎣ ω2 ⎦ ω3 0 r

210


is an integral element of (I, Ω+ ) and every such integral element is of this form. Thus, V4 (I, Ω+ ) = M (1) is diffeomorphic to M ×R3 (where we use p, q, r as coordinates on the R3 -factor) and, on M (1), I (1) is spanned by the 1-forms ϑ1 , ϑ2 , ϑ3, ϑ4 , and ϑ5 . The structure equations of I (1) can now be calculated to be dϑ1 ≡ dϑ2 ≡ 0

(25a) ⎤ ⎡ 0 dϑ3 ⎣ dϑ4 ⎦ ≡ − ⎣ 0 dϑ5 −π6 ⎡ (25b)

0 π6 0

⎤ ⎤ ⎡ ⎤ ⎡ π6 ω1 2ap ω1 ∧ ω2 ⎦ π 7 ⎦ ∧ ⎣ ω2 ⎦ + ⎣ 0 0 π8 ω3

where all of the congruences are taken mod I (1) and π6 , π7 , and π8 restrict to each fiber of M (1) → M to become dp, dq, and dr respectively. Note that (25b) implies that the locus V4 (I (1), Ω+ ) ⊂ V4 (M (1) ) lies entirely over the locus in M (1) defined by the equation ap = 0. In fact, if we let A ⊂ M (1) denote the submanifold defined by the equation a = 0 and let P ⊂ M (1) denote the submanifold defined by the equation p = 0, then V4 (I (1), Ω+ ) can be written as the (non-disjoint) union A(1) ∪ P (1) where A(1) ∼ = A × R3 (respectively, P (1) ∼ = P × R3 ) (1) (1) is the submanifold of V4 (M ) consisting of those elements of V4 (I , Ω+ ) whose base point lies in A (resp., P ) and (using coordinates t, u, v on R3 ) such that the corresponding integral element is annihilated by the 1-forms ϑ1 , ϑ2 , ϑ3 , ϑ4, ϑ5 and the 1-forms ⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ϑ6 π6 0 0 t ω1 ⎣ ϑ 7 ⎦ = ⎣ π 7 ⎦ − ⎣ 0 t u ⎦ ⎣ ω2 ⎦ . (26) ϑ8 π8 ω3 −t 0 v Thus, V4 (I (1), Ω+ ) is singular, being the union of two smooth manifolds of dimension 14 in G4 (M (1)) which intersect transversely along a submanifold diffeomorphic to (A ∩ P ) × R3 of dimension 13. According to the stratification procedure outlined at the beginning of this section, we should regard each of the strata A(1) ∩ P (1), A(1) \(A(1) ∩ P (1)), and P (1)\(A(1) ∩ P (1)) as separate manifolds on which to define the systems I (2). Let us begin with the stratum (A\(A∩P ))×R3 = A(1)\(A(1) ∩P (1)) in M (2). On A\(A ∩ P ) ⊂ M (1), we have a a = 0 and p = 0. Looking back at the formulas (23) and (24), we see that, restricted to A\(A ∩ P ), we have ϑ3 = −pω3 . In particular, ω3 ≡ 0 mod I (1) , so there cannot be any integral elements of (I (2) , Ω+ ) at points of A(1) \(A(1) ∩ P (1)) ⊂ M (2). Thus, there are no integral manifolds of (I, Ω+ ) whose canonical lifts lie in A\(A ∩ P ) ⊂ M (1) . It follows that we may ignore the stratum A(1) \(A(1) ∩ P (1)) in M (2) for the remainder of the discussion. The remaining two strata fit together to be the smooth submanifold P (1) ⊂ V4 (I (1), Ω+ ). For simplicity, we shall therefore let (I (2), Ω+ ) denote the restriction of the canonical differential system on V4 (M (1)) to P (1) ⊂ M (2) . Since p = 0 on P (1), a calculation gives that, on P (1) π6 = 2a((ac − q)ω1 − (ab + r)ω2 )


211

and hence that ϑ6 = 2a(ac − q)ω1 − 2a(ab + r)ω2 − tω3 . Thus, away from the locus Z defined by the equations 2a(ac − q) = 2a(ab + r) = t = 0, we see that no integral element of I (2) can satisfy the independence condition. This locus Z is the union of two smooth submanifolds: Z1 which is the locus defined by the equations p = t = a = 0 and Z2 which is the locus defined by the equations p = t = q − ac = r + ab = 0. First, consider the locus Z1 ∼ = (A ∩ P ) × R2 where we use u, v as coordinates on 2 the R -factor. If we restrict the system I (1) to A∩ P ⊂ M (1), then we may compute that ϑ3 = π6 = 0. The structure equations for I (1) restricted to A ∩ P may now be computed to be dϑ1 ≡ dϑ2 ≡ 0 (27)

dϑ4 ≡ −π7 ∧ ω3

mod I (1)

dϑ5 ≡ −π8 ∧ ω3 . The reduced Cartan character sequence for (I (1), Ω+ ) restricted to A ∩ P is then clearly (s0 , s1 , s2 , s3 , s4 ) = (4, 2, 0, 0, 0). Moreover, it is clear that there does exist a 2-parameter family of integral elements of (I (1) , Ω+ ) at each point of A ∩ P . In fact, Z1 = V4 (I (1) , Ω+ ) when the underlying manifold is A ∩ P . Thus, by Cartan’s Test, (I (1) , Ω+ ) is involutive on A ∩ P . From this it is easy to see that (I (2), Ω+ ) is involutive on Z1 . (A direct proof is easy in this case, but see the next section.) It is interesting to note that if we let Z0 ⊂ M denote the locus defined by the equation a = 0, then (I, Ω+ ) is actually involutive (with Cartan character sequence (s0 , s1 , s2 , s3 , s4 ) = (2, 2, 0, 0, 0)) when restricted to Z0 . Moreover, when (I, Ω+ ) is restricted to Z0 , we get V4 (I, Ω+ ) = A ∩ P . Next, consider the locus Z2 . It is straightforward to calculate that, restricted to Z2 , we have π6 = π7 = π8 = 0. It follows from (26) that there are no integral elements of (I (2) , Ω+ ) restricted to Z2 except along the sublocus Z3 ⊂ Z2 defined by the additional conditions u = v = 0. When we restrict to Z3 , then the forms ϑ6 , ϑ7 , and ϑ8 all vanish and the structure equations (26) imply that the remaining 1-forms ϑ1 , ϑ2, ϑ3 , ϑ4 , and ϑ5 in I (2) form a Frobenius system. Note that Z3 is actually diffeomorphic to M via its natural projection to M . Thus, by prolongation, we arrive at the following classification of the framings which correspond to Riemannian submersions: There are two types of such framings. The first type consists of the integrals of the involutive system (I, Ω+ ) on F ×R2 where I is generated by the rank 2 Pfaffian system I which is generated by the two 1-forms (28)

ϑ1 = ω13 − bω3 ϑ2 = ω23 − cω3 .

The Cartan character sequence of (I, Ω+) is (s0 , s1 , s2 , s3 , s4 ) = (2, 2, 0, 0, 0). Thus, the general solutions depend on two functions of one variable.

212


In fact, it is quite easy to describe these solutions geometrically using the Cauchy characteristic foliation of the system I defined by (28). The reader may want to verify the following description of the corresponding Riemannian submersions: Note that the totally geodesic surfaces in N 3 depend on 3 parameters. Let Γ3 denote this space. Any curve γ in Γ which is in “general position” represents a 1-parameter family of totally geodesic surfaces in N 3 which foliates an open set U ⊂ N . Let Fγ denote this foliation on U . Let Lγ denote the orthogonal foliation of U by curves. Then the projection U → U/Lγ is a Riemannian submersion where the quotient metric is such that each of the leaves of Fγ projects isometrically onto U/Lγ . Of course, in this local description, we are ignoring all the difficulties caused by the (possible) non-Hausdorf nature of the quotient. The second type corresponds to the integrals of the Frobenius system I on F ×R3 generated by the 1-forms ϑ1 = ω13 (29)

ϑ2 ϑ3 ϑ4 ϑ5

−aω2

−bω3

= ω23 +aω1 −cω3 = da − 2a(bω1 + cω2 ) = db + cω12 +(a2 − K)ω1 −b(bω1 + cω2 ) −acω3 = dc − bω12 +(a2 − K)ω2 −c(bω1 + cω2 ) +abω3 .

We leave the geometric analysis of these integrals and the corresponding Riemannian submersions as an interesting exercise for the reader. §2. Ordinary Prolongation. In this section, we examine the effect that prolongation has when applied to a component Z ⊂ Vn (I) consisting of ordinary integral elements of a differential system I. For convenience, we shall assume that I is generated in positive degree. The following result is due to Matsushima [1953]. Theorem 2.1. Let I ⊂ Ω∗ (M ) be a differential ideal which is generated in positive degree (i.e., I contains no non-zero functions). Let Z ⊂ Vn (I) be a connected component of the space of ordinary integral elements of I. Let (s0 , . . . , sn ) be the sequence of Cartan characters of Z. Let (I (1), Ω) be the restriction to Z of the canonical Pfaffian differential system with independence condition on Gn (T M ). Then (I (1), Ω) is linear and is involutive on Z. Moreover, the sequence of Cartan (1) characters of Vn (I (1), Ω) is given by sp = sp + sp+1 + · · · + sn for all 0 ≤ p ≤ n. Proof. Without loss of generality, we may assume that I contains all forms on M of degree n + 1 or greater. (If not, enlarging I by adjoining all such forms will not affect our hypotheses on Z nor will it affect the sequence of Cartan characters of Z.) Thus, I has no integral elements of dimension larger than n and H(E) = E for E ∈ Z. Let s = dim M − n. To avoid trivialities, we shall assume that s is positive. As usual, we let cp = s0 + · · · + sp for 0 ≤ p ≤ n, and set c−1 = 0 for convenience. Note that cn = s0 +· · · +sn = s is the rank of the polar equations of any integral element E ∈ Z. Recall from Chapter III that Z is a smooth submanifold of Gn (T M ) of codimension c0 + c1 + · · · + cn−1 = ns0 + (n − 1)s1 + · · · + sn−1 = ns − (s1 + 2s2 + · · · + nsn ).

§2. Ordinary Prolongation

213

The conclusions of the theorem are local statements about the system I (1), so it suffices to examine the differential ideal I (1) in a neighborhood of an arbitrary element of Z. Let E be an element of Z and let z ∈ M be its base point. Then there exists a local coordinate system x1 , . . . , xn , y1 , . . . , ys centered at z on a zneighborhood U ⊂ M so that E is spanned by the vectors {∂/∂xi | 1 ≤ i ≤ n} at z. Moreover, we may assume that, for all p ≤ n, the subspace Ep ⊂ E spanned by the vectors {∂/∂xi | 1 ≤ i ≤ p} at z is a regular p-dimensional element of I and that the polar space H(Ep ) is spanned by the vectors {∂/∂xi | 1 ≤ i ≤ n} at z together with the vectors {∂/∂ya | a > cp } at z. In particular, the polar equations E(Ep ) are spanned by the 1-forms {dya | a ≤ cp } at z. Let Ω = dx1 ∧ dx2 ∧ · · · ∧ dxn . As is our usual convention, let Gn (T U, Ω) denote the space of n-planes in Gn (T U ) on which Ω restricts to be non-zero. We define the functions pai on Gn (T U, Ω) as usual so that E˜ ∈ Gn (T U, Ω) is annihilated by i ˜ . The functions (x, y, p) then form a coordinate system the 1-forms dya − pai (E)dx on Gn (T U, Ω) centered at E. Moreover, the 1-forms ϑa = dya − pai dxi span the canonical Pfaffian system on Gn (T U, Ω) ⊂ Gn (T M ). Also, in accordance with our ˜ ∈ Gn (T U, Ω) which is based at w ∈ U , we let earlier notation, for each E (30)

a ˜ = (∂/∂xi + pa (E)∂/∂y ˜ Xi (E) )|w . i

denote the basis of E˜ dual to the 1-forms dx1 , dx2, . . . , dxn. Let π : Gn (T U, Ω) → U be the base-point projection. Then for every exterior form ϕ on U which is of degree p + 1 ≤ n, the corresponding (p + 1)-form π ∗ (ϕ) has a unique expansion on Gn (T U, Ω) of the form (31)

π ∗ (ϕ) = 1/(p + 1)!

FK dxK + 1/p!

fbJ ϑb ∧ dxJ + Q.

In (31), the summation in the first term is over all (skew-symmetric) multi-indices K from the range 1, . . . , n and of degree p + 1, the summation in the second term is over (skew-symmetric) multi-indices J from the same range but of degree p and over all b in the range 1, . . . , s, while the last term Q is a form of degree p + 1 which is at least quadratic in the terms {ϑb | 1 ≤ b ≤ s}. It is elementary that the functions FK on Gn (T U, Ω) satisfy (32)

˜ = ϕ(Xk0 (E), ˜ Xk1 (E), ˜ . . . , Xkp (E)) ˜ FK (E)

for each multi-index K = (k0 , k1, . . . , kp). To get corresponding formulae for the functions fbJ , we compute the exterior derivative of both sides of (31) and reduce modulo the ideal generated by the contact forms {ϑb }. This gives the formula (33)

dFK ≡ 1/p!

fbJ dpbi

mod {ϑ, dx}.

b,iJ=K

Now, recall, from Chapter III, our convention which defined the level, λ(a), of an integer a in the range 1 ≤ a ≤ s to be the integer k so that ck−1 < a ≤ ck . We let P = {(i, a) | 1 ≤ i ≤ λ(a)} denote the set of principal pairs of indices. Since there are sk integers a in the range 1, . . . , s satisfying λ(a) = k, it follows that P contains s1 + 2s2 + · · · + nsn pairs of indices. Any pair (j, a) satisfying λ(a) < j ≤ n will be referred to as non-principal.

214


Let ϕ1 , . . . , ϕs be a polar sequence (see Chapter III) for the integral flag (0) ⊂ E1 ⊂ E2 ⊂ · · · ⊂ En = E. Thus, ϕa is a form in I of degree λ(a) + 1. By choosing our polar sequence appropriately, we may even suppose that, for all a with λ(a) = 0, we have ϕa |z = dya while, for all a with λ(a) > 0, we have ϕa (v, ∂/∂x1 , . . . , ∂/∂xλ(a)) = dya (v)

(34) for all v ∈ Tz M . Let (35)

π ∗ (ϕa ) = 1/(p + 1)!

a FK dxK + 1/p!

a b fbJ ϑ ∧ dxJ + Qa

a be the expansion of π ∗ (ϕa ) as in (31). Then by (32), the functions FK vanish identically on Z ∩ Gn (T U, Ω) for all a and all multi-indices K of degree λ(a) + 1. a It is easy to show that the relations (34) imply that fb12...λ(a) (E) = δba (Kronecker δ) for all a and b. It then follows from (33) that, at E, for each non-principal pair (j, a) we have a dFj12...λ(a) ≡ dpaj

(36)

mod {ϑ, dx, {dpbi}i≤λ(a)}.

a Thus, the collection of functions F = {Fj12...λ(a) | (j, a) non-principal} has linearly independent differentials on a neighborhood of E in Gn (T U, Ω). Since the locus of common zeroes of F contains Z ∩ Gn (T U, Ω) by construction and since the codimension of Z in Gn (T U, Ω) is equal to ns − (s1 + 2s2 + · · · + nsn ) = the number of non-principal pairs, there exists an open neighborhood W of E ∈ Gn (T U, Ω) so that the functions in F have linearly independent differentials on W and so that the set Z ∩ W is the common set of zeroes of these functions in W . a Moreover, on Z ∩ W the equations Fj12...λ(a) = 0 imply

a 0 = dFj12...λ(a) ≡ 1/p!

(37) ≡

a fbJ dpbi mod {ϑ, dx}

b,iJ=j12...λ(a) a fb12...λ(a) dpbj mod {ϑ, dx, {dpbi}i≤λ(a) }.

b a Combining this with the fact that fb12...λ(a) (E) = δba , it follows that, by shrinking W if necessary, we may suppose that the ns − (s1 + 2s2 + · · · + nsn ) relations (37) may be expressed in the form

(38)

dpaj ≡

ai Bjb dpbi

mod {ϑ, dx}

i≤min(λ(b),j)

where (j, a) is non-principal. In particular, on Z ∩ W , for each non-principal pair (j, a), the function paj can be expressed as a function of the variables x, y, and {pbi | i ≤ min(λ(b), j)}. Thus, the functions x, y, {paj | j ≤ λ(a)} form a coordinate system on Z ∩ W centered at E and the n + s + s1 + 2s2 + · · · + nsn 1-forms {dxi | 1 ≤ i ≤ n}, {ϑa | 1 ≤ a ≤ s}, and {dpai | i ≤ λ(a)} are a coframing on Z ∩ W .

§2. Ordinary Prolongation

215

Finally, we may suppose, by shrinking W if necessary, that, for any E˜ ∈ Z ∩ W , ˜1 ⊂ E˜2 ⊂ · · · ⊂ E ˜ n = E, ˜ defined by letting E˜p be the span the integral flag (0) ⊂ E ˜ ˜ of the vectors X1 (E), . . . , Xp (E), is a regular flag. Let I (1) be the Pfaffian system on Z ∩ W generated by the 1-forms b {ϑ | 1 ≤ b ≤ s}. Then I (1) generates the differential ideal I (1) restricted to Z ∩ W . In order to prove involutivity of (I (1) , Ω), we need to compute the expressions {dϑb | 1 ≤ b ≤ s} modulo I (1) . Now, on Z ∩ W , we have dϑa = −dpai ∧ dxi .

(39)

The equations (38, 39) constitute the structure equations of the system I (1) . Using the reduced flag determined by the sequence (dx1 , dx2, . . . , dxn), we see by (38) that the reduced characters of I (1) are given for 0 ≤ p ≤ n by s˜p = sp + sp+1 + · · · + sn .

(40)

Now, let π ¯ ia = dpai for all principal pairs (i, a) and define ai b Bjb π ¯i (41) π ¯ja = i≤min(λ(b),j)

for every non-principal pair (j, a). Then by (38) we have structure equations for I (1) of the form (42) (43)

1 πia ∧ dxi + Tija dxi ∧ dxj mod I (1) dϑa ≡ −¯ 2 a ai b Bjb π ¯i mod I (1) . π ¯j ≡ i≤min(λ(b),j)

In order to prove that the torsion of this system vanishes, we need to prove the existence of functions Laij on Z ∩ W so that the equations (44)

Lajk +

Laij − Laji = Tija

ai b i≤min(λ(b),j) Bjb Lik

= 0 (λ(a) < j)

hold. In order to prove that the symbol relations of I (1) are involutive with Cartan characters given by (40), we need to show that the space of solutions of the homogeneous equations associated to (44) is of dimension s˜1 + 2˜ s2 + · · · + n˜ sn = 1 1 s1 + 3s2 + · · · + 2 p(p + 1)sp + · · · + 2 n(n + 1)sn at each point of Z ∩ W . (Note that Cartan’s inequality already tells us that the homogeneous solution space cannot have dimension larger than this number.) We will now prove both of these assertions together. Let β : Z ∩ W → U be the restriction of the base-point projection π to Z ∩ W . Note that for any ϕ ∈ I, the functions FK in the expansion (35) of β ∗ (ϕ) must all be zero. Thus, β ∗ (ϕ) ≡ 0 mod I (1) for all ϕ ∈ I. In particular, the expansion (35) simplifies now to a b (45) β ∗ (ϕa ) = 1/p! fbJ ϑ ∧ dxJ + Qa

216


where the forms Qa are of degree at least 2 in the terms {ϑb | 1 ≤ b ≤ s} and the a {fbJ | deg J = λ(a)} are functions defined on Z ∩ W . Note that on Z ∩ W we also have a a (46) 0 = dFK ≡ 1/p! fbJ dpbi mod {ϑ, dx}, b,iJ=K

which implies for all a and K that a b fbJ π ¯i (47) 0=

mod {ϑ, dx}.

b,iJ=K

Of course, these relations are simply linear combinations of the relations (41). Indeed, the relations (41) are also linear combinations of these relations. Now, since I is a differential ideal, we have dϕa ∈ I, so it follows that β ∗ (dϕa ) ≡ 0 mod I (1) . Since Qa is at least quadratic in the generators of I (1) , it follows that dQa ≡ 0 mod I (1) . Thus, computing the exterior derivative of (45) and reducing mod I (1) , we get the formula a (48) fbJ dϑb ∧ dxJ ≡ 0 mod I (1) . Substituting the equation (42) into (48) and making use of the equation (47), we get a (49) fbJ Tijb dxi ∧ dxj ∧ dxJ ≡ 0 mod I (1) . ˜ ⊂ Rs ⊗ Λ2 (Rn ) denote the vector space Now, for each E˜ ∈ Z ∩ W , let W(E) a a which consists of the solutions τ = (τij ) of the linear equations τija = −τji and (50)

a ˜ b fbJ (E)τij dxi ∧ dxj ∧ dxJ = 0.

(This sum extends over all b, J (of degree λ(a)), i, and j. Of course, a is fixed.) We claim that the dimension of W(E) is at most D=

n

sp

p=0

n n−p − . 2 2

To see this, note that when we set E˜ = E, then for any triple (i, j, a) with λ(a) < i < j, the coefficient of dxij12...λ(a) in the a’th equation of (50) is given by 2τija + b (terms involving τkl where min(k, l) ≤ λ(b)). It follows at once that there are at least as many linearly independent equations in (50) as there are triples (i, j, a) with λ(a) < i < j. This verifies our upper estimate for the dimension of W(E). Of course, by shrinking W if necessary, we may assume that D is also an upper ˜ for all E˜ ∈ Z ∩ W . bound for the dimension of W(E) ˜ ⊂ Rs ⊗ Rn ⊗ Rn denote the vector space Now, for each E˜ ∈ Z ∩ W , let L(E) a which consists of the solutions l = (lij ) of the linear equations (51)

b,iJ=K

a ˜ b fbJ (E)lij = 0.

§3. The Prolongation Theorem

217

Because of our remarks following (47) above, we know that these equations are equivalent to the equations

a + ljk

(52)

ai ˜ b Bjb (E)lik = 0 (λ(a) < j).

i≤min(λ(b),j)

˜ is n(s1 + 2s2 + · · · + nsn ). It follows that the dimension of L(E) ˜ let δ(l) ∈ Rs ⊗ Λ2 (Rn ) be given by the formula Finally, for each l ∈ L(E), a a δ(l)aij = lij − lji .

(53)

˜ ⊂ W(E). ˜ Moreover, as we It is a consequence of our definitions, that δ(L(E)) have already noted after (44), the kernel of δ cannot have dimension greater than s1 + 3s2 + · · · + 12 p(p + 1)sp + · · · + 12 n(n + 1)sn . It follows that the dimension of ˜ must be at least δ(L(E)) n p=0

sp np −

p+1 2

=

n p=0

n n−p − = D. sp 2 2

˜ ⊂ W(E) ˜ and dim W(E) ˜ ≤ D, it follows that we must have both Since δ(L(E)) (54)

˜ = W(E) ˜ δ(L(E))

and (55)

dim ker δ =

n p+1 p=0

2

sp

˜ ∈ Z ∩ W . Thus, L and W are smooth vector bundles over Z ∩ W and the for all E map δ : L → W is a smooth surjection. Since by (49), T is a smooth section of W ˜ ˜ = (T a (E)), it follows that there exists a smooth section L = (Laij ) of where T (E) ij L so that T = δ(L). This completes the verification of the existence of a solution to the equation (44) and the verification of the required dimension of the space of solutions to the homogeneous equations (which is the rank of the bundle ker δ.) By Cartan’s Test, it follows that (I (1) , Ω) is involutive. §3. The Prolongation Theorem. In this section, we prove a version of the Cartan–Kuranishi prolongation theorem. The aim of this theorem is to reduce the problem of finding the integrals of a differential system to that of finding the integrals of a differential system which is in involution. While the actual theorem we prove is not quite strong enough to do this, it suffices for the analysis of most differential systems which arise in practice. We begin with the following fundamental definition. Definition 3.1. Given an exterior differential ideal I ⊂ Ω∗ (M ) and an integer n, a prolongation sequence for I is a sequence of manifolds {Mk | k ≥ 0} (where

218


M0 = M ) together with immersions ιk : Mk → Gn (T Mk−1 ) for k > 0 with the following properties: (i) The map ι¯k : Mk → Mk−1 is a submersion for all k > 0. Here, ¯ιk is the composition Mk → Gn (T Mk−1 ) → Mk−1 . (ii) ι1 (M1 ) ⊂ Gn (I) and for all k ≥ 1, ιk+1 (Mk+1 ) ⊂ Gn (I (k) , Ω(k)) where (k) (I , Ω(k)) is the pull-back to Mk of the canonical differential system with independence condition on Gn (T Mk−1 ). To define a prolongation sequence for a pair (I, Ω), we modify (ii) to require that ι1 (M1 ) ⊂ Gn (I, Ω). Note that one particular example of a prolongation sequence when I = 0 is to let M1 = Gn (T M ), let (I (1) , Ω(1)) be the canonical exterior differential system with independence condition on Gn (T M ), and then, by induction, define Mk+1 = Gn (I (k), Ω(k)) and let (I (k+1), Ω(k+1)) be the restriction to Mk+1 of the canonical exterior differential system with independence condition on Gn (T Mk ) for all k > 0. The resulting sequence is called the tautological prolongation sequence. We can now state our main theorem concerning prolongation sequences. Theorem 3.2. If S = {(Mk , ιk ) | k > 0} is a prolongation sequence for I over M = M0 , then there exists an integer k0 such that, for k ≥ k0 , each of the systems (I (k), Ω(k)) is involutive and moreover, ιk+1 (Mk+1 ) is an open subset of Gn (I (k), Ω(k)). Before we begin the proof of Theorem 3.2, we shall establish a piece of notation concerning prolongation sequences that will be useful in the sequel. Fix a prolongation sequence S = {(Mk , ιk ) | k > 0} over a base manifold M = M0 . A sequence of elements y = (y0 , y1 , y2 , . . . ) with yk ∈ Mk which satisfies the condition ι¯k (yk ) = yk−1 for all k > 0 will be called a coherent sequence. Note that y1 ⊂ Ty0 M is an n-plane by definition. Let us define Qy = Ty0 M/y1 and Ey = y1 . Also, we remind the reader of the following algebraic notation from Chapters IV and V (also, see Chapter VIII) which will be used extensively in the proof. If V is a (real) vector space of dimension n, then there is a natural pairing S k (V ∗ ) ⊗ V → S k−1 (V ∗ ) which, in the interpretation of S k (V ∗ ) as the space of polynomial functions on V of degree k, corresponds to partial differentiation. The extension of this mapping, by tensoring with a space W , W ⊗ S k (V ∗ ) ⊗ V → W ⊗ S k−1 (V ∗ ) is the obvious one. Moreover, if α and β are differential forms on a manifold M which have values in the spaces W ⊗ S k (V ∗ ) and V respectively, then α ∧ β (or simply αβ if α is of degree 0) will denote the W ⊗ S k−1 (V ∗ )-valued differential form obtained by using exterior form multiplication and the above pairing. One algebraic lemma which we shall use rather frequently is the following consequence of the polynomial version of Poincaré’s lemma: If β is a V -valued 1-form on M whose components are linearly independent and a is a function on M with values in W ⊗ S k (V ∗ ) ⊗ V ∗ with the property that (aβ) ∧ β = 0, then a actually has values in W ⊗ S k+1 (V ∗ ) ⊂ W ⊗ S k (V ∗ ) ⊗ V ∗ . Proof of Theorem 3.2. Fix a coherent sequence y in S. Let dim M = n + s. To avoid trivialities, we assume that n and s are both positive. Let U0 be an open neighborhood of y0 on which there exist local coordinates (x1 , . . . , xn , u1 , . . . , us ) = (x, u) centered on y0 so that Ω(0) = dx1 ∧ dx2 ∧ · · · ∧ dxn does not vanish on y1 . Clearly, du|y0 induces an isomorphism of Qy with Rs , and dx|y0 induces an

§3. The Prolongation Theorem

219

isomorphism of Ey with Rn . We shall identify these spaces from now on and speak of du as having values in Qy and dx as having values in Ey . (0) Let U1 ⊂ ¯ι−1 1 (U0 ) be an open neighborhood of y1 with the property that Ω does not vanish on any of the n-planes in ι1 (U1 ) ⊂ Gn (T U0 ). There exists a unique function p1 : U1 → Hom(Ey , Qy ) ∼ = Qy ⊗Ey∗ with the property that, for all z1 ∈ U1 , ι1 (z1 ) is the null space of the Qy -valued 1-form du−p1 (z1 )dx at z0 = ¯ι1 (z1 ). By our hypothesis that ι1 be an immersion, it follows that (¯ι1 , p1 ) : U1 → U0 × Qy ⊗ Ey∗ is an immersion. In particular, p1 is an immersion when restricted to any fiber of ¯ι1 . It follows that, for each z1 ∈ U1 , dp1 induces an injection ker(d¯ι1 )|z1 → Qy ⊗ Ey∗ . We let A(1) (z1 ) ⊂ Qy ⊗ Ey∗ denote dp1 (ker(d¯ι1 )|z1 ). Then A(1) is a smooth subbundle of the trivial bundle over U1 whose fiber is Qy ⊗ Ey∗ . By definition, dp1 induces an isomorphism of the sub-bundle ker(d¯ι1 ) ⊂ T U1 with the bundle A(1) . For convenience of notation, we define A(0) to be the trivial bundle over U0 whose typical fiber is Qy . By pull-back, we regard A(0) as being well-defined over U1 as well. To keep our notation as simple as possible, we shall write x and u for the functions ¯ι∗1 (x) and ι¯∗1 (u) on U1 . Then the components of the Qy -valued 1-form ϑ0 = du − p1 dx span the pull-back of the canonical Pfaffian system I (1) on Gn (T U0 ) to U1 . Moreover, the canonical independence condition may be taken to be Ω(1) = dx1 ∧ dx2 ∧ · · · ∧ dxn . We now want to derive the structure equations of (I (1), Ω(1)). Let σ1 : U1 → M2 be a section of the submersion ¯ι2 which satisfies σ1 (y1 ) = y2 . (We may have to shrink U1 to do this.) Let P (z1 ) = ι2 (σ1 (z1 )). Then P ⊂ T U1 is a rank n sub-bundle whose fiber at each point of U1 is an integral element of (I (1), Ω(1)). It follows easily that there exists a unique Qy ⊗ Ey∗ -valued 1-form π1 on U1 with the following properties: (1) dp1 = π1 + Bϑ0 + T dx on U1 for some B and T . (2) At each z1 ∈ U1 , π1 takes values in A(1)(z1 ). (3) At each z1 ∈ U1 , P (z1 ) is the kernel of π1 and ϑ0 . Note that on ker(d¯ι1 ), we have dp1 = π1 . Also note that the functions B and T have values in Qy ⊗ Ey∗ ⊗ Q∗y and Qy ⊗ Ey∗ ⊗ Ey∗ , respectively. We now have the structure equations (56)

dϑ0 = −dp1 ∧ dx ≡ −(π1 + T dx) ∧ dx mod I (1) .

Moreover, since the distribution P (which is annihilated by π1 and ϑ0 ) is a distribution of integral elements of (I (1), Ω(1)), it follows that (T dx) ∧ dx = 0 on U1 . Thus, we get the structure equation (57)

dϑ0 ≡ −π1 ∧ dx mod I (1) .

Note that the Ey ⊕ Qy ⊕ Qy ⊗ Ey∗ -valued 1-form (dx, ϑ0, π1 ) has constant rank and induces an isomorphism of T U1 with the bundle Ey ⊕ A(0) ⊕ A(1) . At this point, we may continue our construction by induction. Suppose that for each integer k in the range 1 ≤ k < q, we have constructed an open neighborhood Uk of yk so that Uk ⊂ ¯ι−1 k (Uk−1 ). Suppose also that for each k in this range, we have constructed a vector bundle A(k) over Uk which is a sub-bundle of the trivial bundle with typical fiber Qy ⊗ S k (Ey∗ ). By the obvious pull-back, we regard forms and bundles defined over Uj for j < k as being well-defined over Uk . We

220


suppose that these bundles satisfy A(k) ⊂ (A(k−1))(1) for all 1 ≤ k < q. (Note that we do not assume that (A(k−1) )(1) is a vector bundle, indeed, it may not have constant rank.) Finally, we suppose that we have constructed a sequence of 1-forms (dx, ϑ0, ϑ1 , . . . , ϑq−2, πq−1 ) with the following properties: (1) For k < q − 1, the Qy ⊗ S k (Ey∗ )-valued 1-form ϑk is well defined on Uk+1 and takes values in the sub-bundle A(k). The Qy ⊗ S q−1 (Ey∗ )-valued 1-form πq−1 is well defined on Uq−1 and takes the values in the sub-bundle A(q−1) . (2) At each point of Uq−1 , the 1-form (dx, ϑ0, ϑ1 , . . . , ϑq−2 , πq−1) induces an isomorphism of T Uq−1 with Ey ⊕ A(0) ⊕ A(1) · · · ⊕ A(q−1) . (3) For each k < q, the components of the forms {ϑ0 , ϑ1, . . . , ϑk−1} span the Pfaffian system I (k) restricted to Uk . Furthermore, they satisfy the structure equations dϑj ≡ −ϑj+1 ∧ dx mod {ϑ0 , ϑ1, . . . , ϑj }

(58) for j < q − 2 and (59)

dϑq−2 ≡ −πq−1 ∧ dx mod {ϑ1 , ϑ1 , . . . , ϑq−2}.

ι−1 Now, let Uq ⊂ ¯ q (Uq−1 ) be an open neighborhood of yq . Then there exists a unique function pq : Uq → Qy ⊗ S q−1 (Ey∗ ) ⊗ Ey∗ with the property that if zq−1 = ¯ιq (zq ) where zq ∈ Uq , then pq (zq ) ∈ A(q−1) (zq−1 ) ⊗ Ey∗ and so that the A(q−1) (zq−1 )-valued 1-form πq−1 − pq (zq )dx annihilates the integral element ιq (zq ) of (I (q−1) , Ω(q−1)). Since πq−1 ∧ dx ∈ I (q−1) must vanish on ιq (zq ), we must have (pq (zq )dx) ∧ dx = 0. By the lemma, this implies that pq (zq ) ∈ Qy ⊗ S q (Ey∗ ). Thus, we must have (60)

pq (zq ) ∈ (A(q−1) (zq−1 ) ⊗ Ey∗ ) ∩ Qy ⊗ S q (Ey∗ ) = (A(q−1) (zq−1 ))(1) .

Just as in the above discussion of the case q = 1, the assumption that ιq is an immersion implies pq is an immersion when restricted to the fibers of ι¯q . We define A(q) (zq ) ⊂ Qy ⊗ S q (Ey∗ ) to be the vector space dpq (ker(d¯ιq )|zq ). It then follows that A(q) is a smooth sub-bundle of the trivial bundle over Uq whose typical fiber is (q−1) (zq−1 ))(1) , it follows Qy ⊗ S q (Ey∗ ). Since, by (60), we have pq (¯ι−1 q (zq−1 )) ⊂ (A (q) (q−1) (1) −1 that A (zq ) ⊂ (A (zq−1 )) for all zq ∈ ¯ιq (zq−1 ). Thus A(q) ⊂ (A(q−1) )(1) . Now define ϑq−1 = πq−1 − pq dx. Then the equation (58) holds for all j < q − 1. Moreover, it is clear that the components of the forms {ϑ0 , ϑ1 , . . . , ϑq−2 , ϑq−1 } generate the Pfaffian system I (q) on Uq . It remains to construct the appropriate form πq with values in A(q) and verify the analogue of (59) to complete the induction step. First, we note that since πq−1 is well-defined on Uq−1 , it follows that dπq−1 can be expressed in terms of the forms {dx, ϑ0, ϑ1 , . . . , ϑq−2 , πq−1 }. Using the fact that πq−1 ≡ pq dx mod I (q) , we then obtain a formula of the form (61)

dπq−1 ≡ Tq−1 (dx ∧ dx) mod I (q)

where Tq−1 is some function on Uq with values in A(q−1) ⊗ Λ2 (Ey∗ ).

§4. The Process of Prolongation

221

It follows that we have the structure equation (62)

dϑq−1 ≡ −dpq ∧ dx + Tq−1 (dx ∧ dx) mod I (q) .

Next let σq : Uq → Mq+1 be a section of the submersion ¯ιq+1 which satisfies σq (yq ) = yq+1 . (We may have to shrink Uq to do this.) Let P (zq ) = ιq+1 (σq (zq )). Then P ⊂ T Uq is a rank n sub-bundle whose value at each point of Uq is an integral element of (I (q) , Ω(q)). It follows easily that there exists a unique Qy ⊗ S q (Ey∗ )valued 1-form πq on Uq with the following properties: (1) dpq ≡ π1 + Rq dx mod I (q) on Uq for some function Rq which has values in Qy ⊗ S q (Ey∗ ) ⊗ Ey∗ . (2) At each zq ∈ Uq , πq takes values in A(q) (zq ). (3) At each zq ∈ Uq , P (zq ) is in the kernel of πq . Note that on ker(d¯ιq ), we have dpq = πq . It follows that the 1-form (dx, ϑ0, ϑ1, . . . , ϑq−1 , πq ) induces an isomorphism of T Uq with Ey ⊕ A(0) ⊕ A(1) · · · ⊕ A(q) . Also, equation (7) becomes (63)

dϑq−1 ≡ −πq ∧ dx − (Rq dx) ∧ dx + Tq−1 (dx ∧ dx) mod I (q) .

However, since the distribution P is a distribution of integral elements of I (q) and since πq vanishes on P , it follows that (Rq dx) ∧ dx − Tq−1 (dx ∧ dx) = 0 on Uq . Thus, (63) simplifies to (64)

dϑq−1 ≡ −πq ∧ dx mod I (q) .

Setting Ω(q) = dx1 ∧ dx2 ∧ · · · ∧ dxn , this completes the induction step. Now, for every coherent sequence z = (z0 , z1 , z2 , . . . ) with zk ∈ Uk , we have the sequence of vector spaces A(k)(z) = A(k) (zk ) ⊂ Qy ⊗ S k (Ey∗ ) which have the property that A(k) (z) ⊂ (A(k−1)(z))(1) for all k > 0 and the property that dim A(k) (z) is independent of the coherent sequence z. By Proposition 3.10 of Chapter VIII, it follows that there exists an integer k0 >> 0 so that, for all k ≥ k0 and all coherent (1) sequences z, A(k) (z) is involutive and moreover A(k+1)(z) = A(k)(z) . Moreover, k0 can be bounded above by a constant which depends only on the sequence of integers dk = dim A(k) (z). Now assume that k ≥ k0 is fixed. The structure equations of (I (k), Ω(k)) on Uk are then given by (65)

dϑj ≡ 0 mod I (k)

0≤ j < k−1

dϑk−1 ≡ −πk ∧ dx mod I (k) .

Since πk is a 1-form which maps ker(d¯ιk ) surjectively onto A(k) ⊂ Qy ⊗ S k (Ey∗ ), and since, by Proposition 3.10 of Chapter VIII, A(k) (zk ) is involutive as a tableau in (A(k−1)(zk )) ⊗ Ey∗ , it follows immediately from the above structure equations (1) that (I (k), Ω(k)) is involutive on Uk . Since A(k+1) (zk+1 ) = A(k)(zk ) for all zk+1 ∈ ¯ ι−1 (z ), it follows for dimension reasons that ι (U ) is an open subset k+1 k+1 k+1 k of Gn (I (k), Ω(k)). Since this construction was undertaken with respect to any coherent sequence y, it follows that there is a k0 sufficiently large so that (I (k), Ω(k)) is involutive on Mk for all k ≥ k0 and that, for dimension reasons, ιk+1 (Mk+1 ) is an open subset of Gn (I (k), Ω(k)) for all such k.

222


§4. The Process of Prolongation. The reader may well wonder about the relevance of Theorem 3.2 for the computation of examples. The aim of prolongation, of course, is to reduce the study of the integral manifolds of an arbitrary differential system to the case of an involutive differential system, the case to which the vast majority of the theory applies. In this last section of the chapter, we will discuss just how successful this program is. Let us begin with the simplest case. Theorem 4.1. Let I ⊂ Ω∗ (M ) be a differential ideal, and let {(M (k), I (k), Ω(k)) | k > 0} be the sequence of its prolongations. Suppose that, for each k > 0, the space Vn (I (k−1), Ω(k−1)) = M (k) is a smooth submanifold of Gn (T M (k−1)) and that the projection M (k) → M (k−1) is a surjective submersion. Then there exists an integer k0 ≥ 0 such that (I (k) , Ω(k)) is involutive on M (k) for all k ≥ k0 . Proof. This follows immediately from Theorem 3.2 since {(M (k), I (k), Ω(k)) | k > 0} is clearly a prolongation sequence. While Theorem 4.1 is somewhat satisfying, it is of limited use in practice for the following reason. In order to verify the hypotheses of Theorem 4.1 for a given differential system I, one must be able to compute the entire prolongation sequence {(M (k), I (k), Ω(k)) | k > 0}. In the process of doing this computation, of course, one usually checks whether (I (k), Ω(k)) is involutive while one is computing M (k+1). Thus, in practice, before one can apply Theorem 4.1 (assuming that it does, indeed, apply), one finds an involutive prolongation of I anyway, and then Theorem 2.1 takes over to ensure that all higher prolongations are involutive. Nevertheless, there are cases where Theorem 4.1 is useful. Although the terminology is not explained until Chapter VIII, we give one such example here because of its close association with Theorem 4.1. The reader may also want to compare Theorem 2.16 of Chapter IX, where the following result is interpreted in the language of jet bundles. Theorem 4.2. Let I ⊂ J ⊂ T ∗ = T ∗ (M ) be a pair of sub-bundles defining a linear Pfaffian system on a manifold M . Set L = J/I and let A ⊂ I ∗ ⊗ L be the tableau bundle of (I, J). Suppose that (i) There is an integral element of (I, J) at every point of M , (ii) A is 2-acyclic at each point of M , i.e., H p,2 (Ax ) = 0 for all x ∈ M and p > 0, and (iii) The subspaces A(k) ⊂ I ∗ ⊗S (k+1) (L) have constant rank on M for all k ≥ 0. Then the hypotheses of Theorem 4.1 are fulfilled for the prolongation sequence {(M (k), I (k), Ω(k)) | k > 0}. In particular, (I (k), Ω(k)) is involutive for all k sufficiently large. The proof of Theorem 4.2 will only be indicated here. The hypotheses (i) and (iii) (for k = 0) guarantee that Vn (I, Ω) = M (1) is a smooth manifold which submerses onto M . The hypothesis (ii) then guarantees that the Pfaffian system (I (1), Ω(1))


223

has all of its torsion absorbable, i.e., that the set Vn (I (1) , Ω(1)) = M (2) surjects onto M (1). Then (iii) (for k = 1) guarantees that M (2) is a smooth submanifold. This then continues indefinitely. The important point is that the hypotheses (ii) and (iii) ensure that M (k) is a smooth submanifold submersing onto M (k−1) for all k > 1. For more details, see §2 of Chapter VIII. It may seem that Theorem 4.2 would only be marginally more useful than Theorem 4.1. However, for an interesting application, the reader may consult Gasqui [1979b]. The essential point is that the hypotheses of Theorem 4.2 are algebraic pointwise conditions on the structure equations of (I, J) and hence are checkable without having to compute the prolongations (which may depend on high derivatives of the original system). Let us say that a linear Pfaffian system (I, J) which satisfies the hypotheses of Theorem 4.2 is 2-acyclic. We then have the following easy corollary of Theorem 4.2, which may be regarded as a generalization of the Cartan–K¨ ahler theorem for linear Pfaffian systems. Corollary 4.3. If (I, J) is a real analytic, 2-acyclic linear Pfaffian system, then there exist real analytic integral manifolds of (I, J). Proof. By Theorem 4.2, some finite prolongation of (I, J) is real analytic and involutive. Now apply the Cartan–K¨ ahler theorem. Looking over the examples from §1, we see that the first two examples had the property that, at each stage, the prolongation M (k) was a smooth submanifold for which the basepoint projection M (k) → M (k−1) was a surjective submersion. (Actually, we only checked this until we reached a value of k for which the system (I (k), Ω(k)) was involutive on M (k), for then Theorem 2.1 implies that all higher prolongations will have this property.) In practice, however, examples such as Example 1.3 are often encountered. In that example, the reader will recall, that when M (1) = V4 (I, Ω+) we had a submersion M (1) → M , but that the basepoint projection V4 (I (1), Ω+ ) → M (1) was neither surjective nor submersive. Nevertheless, we were able to reduce the analysis of Example 1.3 to the involutive case by applying a sequence of prolongations. It is natural to ask if this can be done for any differential system I. Practically nothing can be said about the prolongations of a general smooth differential system without making various constant rank assumptions which quickly become too cumbersome to be useful. Therefore, for the remainder of this section we shall assume that the exterior differential system I ⊂ Ω∗ (M ) is real analytic with respect to some fixed real analytic structure on M . Then the set Vn (I) is a real analytic subset of Gn (M ) and, as such, has a canonical coarsest real-analytic stratification Vn (I) = Sβ , β∈B

for which each stratum Sβ is a smooth, analytically irreducible submanifold of Gn (M ). Just as in §1, we define the (first) prolongation of I to be the exterior differential system ) ( I (1) , Ω(1) = P|Sβ , Ψ|Sβ , β∈B

where the underlying manifold M (1) is defined to be the disjoint union of the strata Sβ and (P, Ψ) is the canonical differential system with independence condition on

224


Gn (M ). (In the case that an independence condition Ω has been specified, we restrict our attention to Gn (M, Ω) ⊂ Gn (M ).) The higher prolongations are then defined inductively: M (k) is the disjoint union of the strata of the canonical stratification of Vn (I (k−1)) and (I (k), Ω(k)) is the differential system with independence condition got by restricting the canonical differential system with independence condition on Gn (M (k−1), Ω(k−1)) to each stratum of Vn (I (k−1)). Let us denote the inclusion mapping of M (k) into Gn (M (k−1), Ω(k−1)) by ιk and, for 0 ≤ j < k, denote the natural projection mapping from M (k) to M (j) by πjk . Theorem 4.4. If I ⊂ Ω∗ (M ) is a real analytic differential system on M and M (k) is empty for some k > 0, then there are no n-dimensional real analytic integral manifolds of I. Proof. Suppose that f : N n → M were an n-dimensional, irreducible, real analytic integral manifold of I. Then Vn (I) is non-empty and N has a natural lifting f (1) : N → Vn (I). Since N is irreducible, it follows that there is a unique stratum of M (1) which intersects f (1) (N ) in an analytic manifold of dimension n. Let N1 ⊂ N denote the inverse image of this stratum under f (1) . Then f (1) : N1 → M (1) is an n-dimensional, irreducible, real analytic integral manifold of (I (1) , Ω(1)). This process can clearly be continued inductively to produce a non-trivial n-dimensional, irreducible, real analytic integral manifold of (I (k), Ω(k)), denoted by f (k) : Nk → M (k), for all k > 0. In particular, it follows that M (k) is non-empty for all k. It is natural to ask whether the contrapositive converse of Theorem 4.4 is true, namely, whether or not the non-emptiness of M (k) for all k > 0 is sufficient to imply that there are non-empty n-dimensional real analytic integral manifolds of I. Unfortunately, a definitive answer to this question does not seem to be available, though, for a related result due to Malgrange and phrased in the language of jets, the reader may consult Theorem 2.2 of Chapter IX. One statement which would imply this converse is the following: Prolongation Conjecture. If I ⊂ Ω∗ (M ) is a real analytic differential system on M and M (k) is non-empty for all k > 0, then there exists a k0 ≥ 0 so that for all k ≥ k0 there exists an analytic subvariety S (k) ⊂ M (k) which intersects each component of M (k) in a (possibly empty) proper analytic sub-variety so that (I (k), Ω(k)) is involutive on M (k) \ S (k) . Moreover, for every real analytic integral manifold f : N n → M of I there exists an open submanifold Nk ⊂ N together with an immersion f (k) : Nk → M (k) \ S (k) which is an ordinary integral manifold of (I (k), Ω(k)) and which satisfies f = π0k ◦ f (k) on Nk ⊂ N . The reader may be surprised by the appearance of the “singular subvariety” S (k) in the above statement. However, it is easy to see that such a singular subvariety can occur in such a way that it will not be removed by any finite prolongation. Such examples are furnished by the theory of non-regular singular points of ordinary differential equations. Thus, consider the differential system (I, Ω) on M = R2 generated by the single 1-form θ = x2 du − u dx and let Ω = dx. The curve u = 0 is clearly an integral manifold of (I, Ω), and it is easy to see that M (k) = R2 for all k > 0. However, the integral elements which lie over the locus x = 0 are not ordinary for any k > 0. The problem is, of course, caused by the fact that the differentials of the maps πjk drop rank along the locus x = 0.


225

If the Prolongation Conjecture were known to be true, then we could conclude that all of the real analytic integrals of a differential system could be constructed by suitable applications of the Cartan–K¨ ahler theorem, a highly desirable situation. However, at present, we can only say that the evidence for the Prolongation Conjecture is rather strong: Although many examples of prolongation have been computed, no counterexamples to the conjecture have ever been found. Moreover, under appropriate non-degeneracy hypotheses, the Prolongation Conjecture has been proved. One version, due to Kuranishi [1957], is known as the Cartan– Kuranishi prolongation theorem. Unfortunately, the non-degeneracy hypotheses in Kuranishi’s theorem are rather difficult to make explicit (and often difficult to check in practice), so we shall refer the reader to Kuranishi’s paper for the precise statement of his result. The reader may also consult Kuranishi [1967] and Chapter IX for versions of this theorem which apply directly to P.D.E. In practical calculations, all of the maps πjk are submersions away from proper analytic sub-varieties for j sufficiently large, and, in that case, the conjecture follows from Theorem 3.2. Nevertheless, a proof of the full Prolongation Conjecture remains an interesting problem. An alternative approach, which avoids the difficulty caused by the fact that the maps πjk need not have constant rank, is to consider another definition of prolongation (which, for clarity’s sake, we shall call fine prolongation). Let I ⊂ Ω∗ (M ) be a real analytic differential system. Then the set Vn (I) has a coarsest real analytic stratification Vn (I) = Sβ , β∈B

(which may be finer than the original stratification defined above) with the property that each stratum Sβ is connected and smooth and moreover that the basepoint projection π01 : Vn (I) → M has constant rank when restricted to each stratum Sβ . Let us define M 1 to be the disjoint union of the strata Sβ , and let (I 1 , Ω 1 ) denote the differential system with independence condition induced on M 1 by its canonical inclusion into Gn (M ). We then continue the construction inductively, except that we require that Vn (I k−1 , Ω k−1) be given the coarsest smooth stratification with connected strata for which all of the mappings {πjk | 0 ≤ j < k} have constant rank when restricted to any stratum. We then define M k to be the disjoint union of these strata and define (I k , Ω k ) to be the differential system with independence condition induced on M k by its inclusion into Gn (M k−1 , Ω k−1). Note that according to this definition, fine prolongation is not purely inductive, i.e., we do not necessarily have that M 2 is equal to (M 1 ) 1 . We continue to denote the inclusion mapping of M k into Gn (M k−1, Ω k−1 ) by ιk and, for 0 ≤ j < k, we continue to denote the natural projection mapping from M k to M j by πjk . We shall call the sequence S(I) = {(M k , I k, Ω k ) | k > 0} the fine prolongation sequence of I, with similar terminology for a differential system with independence condition (I, Ω). As in §3, we shall call a sequence k (yk ) = yk−1 y = (y0 , y1 , y2 , . . . ) with y0 ∈ M and yk ∈ M k which satisfies πk−1 for all k > 0 a fine coherent sequence for I. The proof of Theorem 4.4 now goes over with only slight modifications to establish the following result.

226


Theorem 4.5. If I ⊂ Ω∗ (M ) is a real analytic differential system on M and there exists an n-dimensional real analytic integral manifold of I, then there exist fine coherent sequences y for I. In particular, M k is non-empty for all k > 0. Proof. As in the proof of Theorem 4.4, given an irreducible real analytic n-dimensional integral manifold of I, f : N n → M , we can construct a decreasing sequence of submanifolds Nk ⊂ N with the property that each Nk is equal to N minus a proper analytic subvariety and so that there exists a lifting f k : Nk → M k of f restricted to Nk which is an integral of (I k , Ω k ). The intersection of all of these submanifolds, N∞ ⊂ N is equal to N minus a countable number of proper real-analytic submanifolds and hence is non-empty. Clearly, if we let y ∈ N∞ be fixed, then the sequence y = (f(y), f 1 (y), f 2 (y), . . . ) is a fine coherent sequence for I. It is certainly reasonable to conjecture that the Prolongation Conjecture is true when “prolongation sequence” is replaced by “fine prolongation sequence,” and it seems that, in the “fine” case, one can even dispense with the singular locus in the statement. However, this version of the Prolongation Conjecture, which might be called the “Fine Prolongation Conjecture,” has not yet been proved either. Finally, let us indicate our reasons for believing a weaker conjecture which is a sort of converse to Theorem 4.5. Conjecture. If I ⊂ Ω∗ (M ) is a real analytic differential system on M and there exists a fine coherent sequence for I, then there exists an n-dimensional real analytic integral manifold of I. The outline of an argument for this conjecture is: Let S(I) be the fine prolongation sequence of I, and let y be a fine coherent sequence for I. For each k > 0, consider the set of linear maps (dπkj )yj : Tyj M j → Tyk M k for all j > k. Because of the identity πji ◦ πkj = πki for all i > j > k, there exists an integer ik > k so that (dπkj )yj (Tyj M j ) = (dπkik )yik (Tyik M ik ) for all j ≥ ik . By the local constancy of the ranks of differentials of the mappings πkj , it follows that there exists a unique,

k smooth, real analytic submanifold My ⊂ M k in a neighborhood of yk with the property that, for all j > k, there is a neighborhood of yj ∈ M j so that πkj re k stricted to this maps into My and is a submersion when regarded as a mapping

k into My . Although the argument is tedious, it seems to be true that the natural

k+1

k inclusion of My into Gn (M k , Ω k ) actually has image in Gn (My , Ω k ), at least on a neighborhood of yk+1 . Assuming this, it then would then follow that

k the sequence {My | k > 1} is a prolongation sequence, as defined in §3. By

k Theorem 3.2, it would follow that the system (My , I k, Ω k ) is involutive for k sufficiently large. In particular, there would exist integral manifolds of such a system, and hence of the original system I. Of course, the Fine Prolongation Conjecture would be a considerably stronger statement. æ

§1. First Order Equations for Two Functions of Two Variables

227

CHAPTER VII EXAMPLES

This chapter is a collection of examples designed to illustrate the various phenomena which occur in the application of differential systems to problems arising in differential geometry and, more generally, in partial differential equations. We have chosen these examples partly on the basis of their intrinsic interest but mainly in the hopes that the reader can use them as a guide to developing facility in computation. §1. First Order Equations for Two Functions of Two Variables. In this example, we shall make a fairly thorough study of the exterior differential systems which arise in the study of systems of first order partial differential equations for two functions of two variables. These cases have the advantage of displaying many of the features of differential systems in general (characteristic variety, torsion, prolongation, etc.) while they remain sufficiently simple that an essentially complete treatment can be undertaken. In the interests of simplicity, we will make constant rank and genericity assumptions whenever they are convenient. If two variables, say z and w, are regarded as functions of two other variables, say x and y, then the general system of r first order partial differential equations for z and w as functions of x and y can be written in the form (1)

F ρ (x, y, z, w, zx, zy , wx, wy ) = 0

1 ≤ ρ ≤ r.

Here the functions F ρ are assumed to be smooth functions of their arguments and, for the sake of simplicity, we assume that at each common zero of the functions F ρ in (x, y, z, w, zx, zy , wx, wy )-space, the equations (1) implicitly define some set of r of the functions zx , zy , wx, wy as smooth functions of the remaining variables. Note that this assumption implies that the number of equations r is at most 4. Examples of such systems of partial differential equations arising in geometry are the volume preserving equation, zx wy − zy wx = 1,

(2) the Cauchy–Riemann equations, (3)

zx − wy = zy + wx = 0,

and the equations which assert that the pair of functions z(x, y) and w(x, y) induce an isometry between the metrics h2 = E(z, w)dz 2 + 2F (z, w)dz ◦ dw + G(z, w)dw 2 and h1 = e(x, y)dx2 + 2f(x, y)dx ◦ dy + g(x, y)dy2 ,

228

VII. Examples

E(z, w)zx2 + 2F (z, w)zxwx+G(z, w)wx2 = e(x, y) (4)

E(z, w)zxzy + F (z, w)(zx wy + zy wx)+G(z, w)wxwy = f(x, y) E(z, w)zy2 + 2F (z, w)zy wy +G(z, w)wy2 = g(x, y).

We want to see how the methods of exterior differential systems apply in such cases. For the sake of uniformity and simplicity of notation, we shall introduce the following change of notation. We rename the variables x, y, z, w, zx, zy , wx, wy as x1 , x2 , z 1 , z 2 , p11 , p12, p21 , p22 respectively. Of course, the reader will recognize these as the standard coordinates on the jet space J 1 (R2 , R2 ). The system of equations (1) then become (5)

F ρ (x1 , x2 , z 1 , z 2 , p11 , p12 , p21 , p22 ) = 0

1 ≤ ρ ≤ r.

By our hypothesis, these equations define a submanifold M ⊂ J 1 (R2 , R2 ) of codimension r such that the source-target projection M → R2 × R2 is a submersion. The contact system on J 1 (R2 , R2 ) is generated by the 1-forms (6) ϑa = dz a − pai dxi 1≤a≤2 and the canonical independence condition is given by the 2-form Ω = dx1 ∧ dx2 . Clearly, the 1-forms {ϑ1 , ϑ2 , dx1 , dx2} remain independent when restricted to M . We let I ⊂ Ω1 (M ) denote the Pfaffian system generated by the pair {ϑ1 , ϑ2 } after restriction to M and let J denote the Pfaffian system generated by the forms {ϑ1 , ϑ2 , dx1, dx2} after restriction to M . The 1-forms {dpai | 1 ≤ i, a ≤ 2} are clearly not linearly independent modulo J after they have been restricted to M . In fact we must have r relations of the form a (7) bρi 1≤ρ≤r a dpi ≡ 0 mod J ρ a where bρi a = ∂F /∂pi . These relations are obtained by expanding the identities on ρ M given by dF = 0. In studying the structure relations (7), it is often helpful to adapt the given bases of I and J to the problem at hand. Thus, on an open set U ⊂ M , let {ϑ1 , ϑ2 , ω1 , ω2 } denote any set of 1-forms which have the property that {ϑ1, ϑ2 } is a basis for the sections of I restricted to U and {ϑ1 , ϑ2, ω1 , ω2 } is a basis for the sections of J restricted to U . Then there exist functions Aab and Bji on U so that the following relations hold:

(8)

ϑa = Aab ϑb dxi ≡ Bji ωj mod I

If we now choose 1-forms πia subject to the condition that a j (9) πia = Aab Bij dpbj + Sij ω mod I a a where Sij = Sji , then we have the familiar structure equations

(10)

dϑa ≡ −

πia ∧ ωi

mod I


together with symbol relations of the form ρ a (11) bρi Cij ωj mod I a πi ≡

229

1≤ρ≤r

where each 2 × 2 matrix bρ = (bρi a ) is obtained from the corresponding matrix ρ −1 ρ −1 b A . bρ = (bρi a ) by the formula b = B For example, in the volume preserving example, the submanifold M is defined by the equation (12)

p11 p22 − p12 p21 = 1.

The symbol relation corresponding to (7) is the equation (13)

p22 dp11 − p21 dp12 − p12 dp21 + p11 dp22 = 0.

If we set ϑa = ϑa and set 1 2 p2 dx = (14) dx2 −p12

−p21 p11

ω1 ω2

a = 0, the relation corresponding to (11) becomes simply then, taking Sij

(15)

π11 + π22 ≡ 0 mod I.

For the purpose of computing Cartan characters and characteristics, relation (15) is much easier to deal with than (13). As another example, consider the isometry problem of the two metrics h1 and h2 mentioned above. Let ω1 , ω2 be an orthonormal coframing for the metric h1 and let η 1 , η 2 be an orthonormal coframing for the metric h2 . Define the matrices A and B so that η a = Aab dz b and dxi = Bji ωj . (Note that A is a function of the z variables and B is a function of the x variables.) If P = (pai ) is the matrix of solutions to the equations given in (4), then one easily sees that the matrix AP B is an orthogonal matrix. Conversely, if we set P = A−1 gB −1 where g is any 2 × 2 orthogonal matrix, then the matrix P satisfies the equations given in (4). Thus, in this case, M is diffeomorphic to R2 × R2 × O(2). Regarding ϑ = (ϑa ), dz = (dz a ), and dx = (dxi ), etc. as columns, we may write on M , (16)

ϑ = g−1 Aϑ = g−1 A(dz − P dx) = g−1 η − ω.

Using the fact that dη = −ϕ∧η and dω = −ψ∧ω where ϕ and ψ are skew-symmetric matrices, we may compute that dϑ = −g−1 ϕ ∧ η + dg−1 ∧ η + ψ ∧ ω (17)

≡ −(g−1 dg + g−1 ϕg − ψ) ∧ ω mod I ≡ −π ∧ ω mod I

where π is a skew-symmetric matrix by virtue of the fact that g is an orthogonal matrix. Thus, in this basis, the symbol relations of the Pfaffian system I are simply (18)

π11 = π22 = π21 + π12 = 0.

230

VII. Examples

The reader should compare these with the symbol relations one gets from (4) by naively differentiating. We will return to this example below. In the general case, we want to take advantage of changes of basis of the form (10)–(11) to reduce the symbol relations to as simple a form as possible. It is clear that we want to normalize the linear span of the matrices {bρ | 1 ≤ ρ ≤ r} in the vector space of all 2 × 2 matrices under the obvious action of the group GL(2, R) × GL(2, R). We will now treat the four possible values of r as separate cases.

Case 1: r = 1. In this case there is a single symbol matrix which we may as well denote by b instead of b1 . The admissible substitutions (10)–(11) allow us to pre- and postmultiply the 2 × 2 matrix b by arbitrary invertible matrices. It follows that the only invariant of the matrix b is its rank, which must be 1 or 2. We will assume that this rank is constant on M . It follows that there are two subcases.

Subcase 1.1: b has rank 1. We may now choose our bases so that the single symbol relation is of the form (19)

π22 ≡ C1 ω1 + C2 ω2

mod I.

Replacing π22 by π22 − C1 ω1 − C2 ω2 and π12 by π12 − C1 ω2 , we may assume that C1 = C2 = 0. Of course, we still have the structure equations (20)

dϑa ≡ −πia ∧ ωi

mod I.

It follows that the torsion of the system vanishes identically. By inspection, we have s1 = 2 and s2 = 1. Moreover, the integral elements at a point depend on s1 + 2s2 = 4 parameters, namely ϑa = 0 and π11 = λ1 ω1 + λ2 ω2 (21)

π21 = λ2 ω1 + λ3 ω2 π12 = λ4 ω1

and of course π22 = 0. Thus, the system is involutive. The r × s symbol matrix σξ at the covector ξ = ξi ωi is, in this case, the 1 × 2 matrix σξ = (0, ξ2 ). Thus, the symbol matrix has rank 1 except when ξ = ξ1 ω1 (when it has rank 0). Thus, the characteristic variety is Ξx = P(Jx /Ix ) ∼ = P1 for all x ∈ M . However, the characteristic sheaf consists of Ξx plus the “embedded component” [ω1 ] ∈ P1 . Examples of this type of equation are given by wy = 0 and wy = z. It is easily shown that the Pfaffian systems I associated to these two equations are not diffeomorphic. (Consider the form ϑ2 in an adapted coframing satisfying π22 = 0 for each of the above systems. This form is canonically defined up to a multiple and yet


231

it is of different Pfaff type for each of the two systems.) Nevertheless, there exists an O.D.E. method for constructing the integral manifolds of such I, thereby avoiding the use of the Cartan–Kähler theorem and proving a smooth existence theorem for the solutions of the given equation. The method is as follows: Construct an adapted coframing ϑ1 , ϑ2 of I with the property that π22 = 0 and that ϑ1 is of Pfaff rank 5. (This can always be done.) Placing ϑ1 in Pfaff normal form (which requires only O.D.E.), we may specify integral manifolds of ϑ1 in terms of a single function of 2 variables. These integral manifolds are of dimension 4 in M . If R is such an integral, then the structure equations show that ϑ2 restricts to be of Pfaff rank 3 on R and hence its integrals can be specified by a single function of 1 variable and have codimension 2 in R. These resulting 2-dimensional integrals are the desired integrals of I. We leave further details to the reader. Subcase 1.2: b has rank 2. This is, in some sense, the generic case for single equations. We may now choose our bases so that the single relation has the form (22)

π21 − π12 = C1 ω1 + C2 ω2

mod I.

Replacing π11 by π11 −C2 ω2 and π22 by π22 +C1 ω1 , we may assume that C1 = C2 = 0. Thus, the torsion always vanishes. We now observe that the system is involutive. By inspection, we see that s1 = 2 and s2 = 1. Moreover, the integral elements at a point of M depend on s1 +2s2 = 4 parameters, namely ϑa = 0 and π11 = λ1 ω1 + λ2 ω2 (23)

π21 = π12 = λ2 ω1 + λ3 ω2 π22 = λ2 ω1 + λ4 ω2 .

The symbol matrix at a covector ξ = ξ1 ω1 + ξ2 ω2 is the 1 × 2 matrix σξ = (−ξ2 , ξ1 ). Thus, the symbol matrix always has rank 1. In particular, it is never injective. It follows that the characteristic variety satisfies Ξx = P(Jx /Ix ) ∼ = P1 for all x ∈ M . All of this is in accordance with the general theory of characteristic varieties developed in Chapter V. Note that the Cartan–Kähler theory predicts that the general solution of such a system will depend on 1 function of 2 variables. The standard example of an equation which falls into this subcase is the equation zy = wx. The general solution of this equation is given by the formula (24)

z = fx and w = fy

where f is an arbitrary function of x and y. A more interesting equation whose “general solution” can be found explicitly is the volume preserving equation (2). The solutions where zx = 0 can be described locally in parametric form by letting h be an arbitrary function of two auxiliary variables s and t which satisfies hst = 0 and setting (25)

x = ht (s, t)

y=t

z = hs (s, t)

w = s.

232

VII. Examples

A similar formula holds for those solutions where zy = 0. Of course, for the generic single equation for two functions of two variables, the general solution cannot be written down so explicitly.

Case 2: r = 2. This is, in many ways, the most interesting of the cases. The symbol matrices {b1 , b2 } span a two-dimensional subspace of the space of 2×2 matrices. We begin by classifying the possible two-dimensional subspaces under the equivalence generated by pre- and post-multiplication by invertible matrices. It turns out that there are exactly 5 equivalence classes. This is proved by noting that the determinant function on the space of 2 × 2 matrices is a conformally invariant quadratic form under the natural action of GL(2, R)×GL(2, R). To see this, note that if R is a 2×2 matrix and (A, B) ∈ GL(2, R) × GL(2, R), then det(ARB −1 ) = (det(A)/det(B))det(R). Thus, a natural invariant of a two-dimensional subspace of the space of 2 × 2 matrices is the type of the quadratic form det after it has been restricted to the given subspace. This crude classification can be refined slightly to give the following list of representatives of the 5 equivalence classes: & 1 x x2 & i B= (26.1) &x ∈ R 0 0 & 1 x 0 & i x B= (26.2) ∈ R & x2 0 & 0 −x1 & i B= (26.3) ∈ R &x x1 x2 & 0 x1 & i B= (26.4) &x ∈ R x2 0 1 & x2 x & i B= (26.5) &x ∈ R x2 −x1 For the sake of simplicity, we shall assume that the symbol relations of our system of two equations have constant type in the above classification. We shall now proceed to analyse each of these subcases separately.

Subcase 2.1: The symbol relations are of type (26.1). In this case, we may change bases so that the symbol relations take the form (27)

π11 ≡ C11 ω1 + C12 ω2 π21 ≡ C21 ω1 + C22 ω2

mod I

Since we may modify the forms πj1 by a symmetric linear combination of the ωj , we see that the torsion of the system vanishes if and only if C12 ≡ C21 . If the torsion of the system does not vanish identically, then we may restrict to the locus C12 − C21 = 0. In the generic case, this gives an extra equation which, adjoined


233

to the given two, gives a system of 3 equations. (We will treat this case below in Case 3.) On the other hand, if the identity C12 ≡ C21 holds, then the structure equations of the system reduce to the form (28)

dϑ1 ≡ 0 2 dϑ ≡ −π12 ∧ ω1 − π22 ∧ ω2

mod I.

By inspection, we have s1 = s2 = 1. The space of integral elements at a point of M clearly depends on 3 parameters. Thus, the system is in involution and the general solution (at least in the analytic case) depends on one function of two variables. The symbol matrix σξ at a non-zero covector ξ = ξi ωi is the 2 × 2 matrix

ξ1 ξ2

0 0

Since σξ always has rank 1, every covector is characteristic. Of course, this implies that a 2-dimensional integral of I cannot be determined by knowledge of any its 1-dimensional subintegrals. Finally, we remark that, in fact, every involutive system of this type is locally equivalent to the standard example (29)

zx = zy = 0.

In particular, the analytic assumption is not needed. This equivalence follows by examining the structure equations (28) closely and showing that they actually “uncouple” into the the equations (30)

dϑ1 ≡ 0 mod ϑ1 dϑ2 ≡ −π12 ∧ ω1 − π22 ∧ ω2 mod ϑ2 .

The result then follows by applying the Frobenius theorem and the Pfaff–Darboux theorem. (See Chapter II.)

Subcase 2.2: The symbol relations are of type (26.2). In this case, we may, by admissible basis change, assume that the symbol relations are of the form π21 ≡ C11 ω1 + C12 ω2 mod I (31) π22 ≡ C21 ω1 + C22 ω2 Replacing π2j by π2j − Cj1 ω1 − Cj2 ω2 and π1j by π1j − Cj1ω2 for j = 1 and 2, we see that we may assume that Cij = 0. Thus, the torsion always vanishes for systems of this type. The structure equations of the system are now of the form (32)

dϑ1 ≡ −π11 ∧ ω1 dϑ2 ≡ −π12 ∧ ω1

mod I.

234

VII. Examples

Inspection now shows that s1 = 2 and s2 = 0. Moreover, the space of integral elements at each point of M is clearly of dimension 2. Thus, the system is involutive. Thus, by the Cartan–K¨ ahler theorem, the integral manifolds (in the real-analytic category) depend on two functions of two variables. The symbol matrix σξ at a non-zero covector ξ = ξ1 ω1 + ξ2 ω2 is the 2 × 2 matrix

ξ2 0

0 ξ2

.

It follows that the characteristic variety at each point x of M is the point [ω1 ] ∈ P(Jx /Ix ). Note that it should be counted with multiplicity 2. An example of this type of P.D.E. is given by the pair of equations zzx + wzy = f(x, y, z, w) (33)

zwx + wwy = g(x, y, z, w)

where, in order to avoid singularities, we assume that the functions f and g do not simultaneously vanish. Actually, using the techniques of Chapter II, an alternative method of describing the integral manifolds of systems in this subcase is available. By the structure equations (32), it follows that the Cartan system of the Pfaffian system I is the Pfaffian system C(I) generated by the 1-forms ϑ1 , ϑ2, ω1 , π11 , π12 . Since M is of dimension 6, it follows that the Cauchy leaves of I are curves. In fact, any two dimensional integral manifold N 2 ⊂ M 6 of (I, Ω) is a union of integral curves of C(I). To see this, note that on any such integral N , the 2-forms π1j ∧ ω1 must vanish. This implies that, on N , there must be relations of the form π1j = λj ω1 for some functions λj . It follows that all of the forms in C(I) vanish when restricted to the integral curves of ω1 on N . Thus, these curves are integral curves of C(I). Conversely, by the general theory of Cauchy characteristics developed in Chapter II, if P 1 is any integral curve of the system (I, ω1 ), then P is transverse to the leaves of C(I). Hence the union of the leaves of C(I) which pass through P is a smooth surface which is an integral of (I, Ω). Thus, the construction of integrals of (I, Ω) reduces to the two O.D.E. problems of constructing the integral curves of C(I) and constructing integrals of (I, ω1 ). In particular, one does not need to apply the Cartan–K¨ ahler theorem. Thus, our description of the integrals of (I, Ω) does not depend on the assumption of real analyticity. In fact, even more is true. Since the system generated by ϑ1 , ϑ2, and ω1 is the restriction to M of the Frobenius system generated by dz 1 , dz 2 , dx1 , and dx2 , it follows that the system generated by ϑ1 , ϑ2 , and ω1 is itself Frobenius. It easily follows that every point of M has local coordinates (a1 , a2 , b1 , b2 , x, y) so that the system I has generators of the form ϑ˜1 = da1 − b1 dx and ϑ˜2 = da2 − b2 dx. (These coordinates can be found using O.D.E. methods alone.) Then the general integral of (I, Ω) in this neighborhood can be described implicitly by the 4 equations aj = f j (x) and bj = df j /dx (j = 1, 2), where f 1 and f 2 are arbitrary functions of x. In particular, all of the differential systems in this subcase are locally equivalent. It is an interesting exercise to reduce the system (33) to this standard form.


235

Subcase 2.3: The symbol relations are of type (26.3). In this case, we may change bases so that the symbol relations take the form π21 − π12 ≡ C11 ω1 + C12 ω2 (34) mod I π22 ≡ C21 ω1 + C22 ω2 Just as before, by adding appropriate linear combinations of the forms ω1 and ω2 to the forms πji , we may assume that the Cij are all zero. Thus, the torsion vanishes. The structure equations now take the form dϑ1 ≡ −π11 ∧ ω1 − π21 ∧ ω2 (35) mod I. dϑ2 ≡ −π21 ∧ ω1 Inspection shows that s1 = 2 and s2 = 0. Moreover, the integral elements of (I, Ω) at a point x of M depend on 2 parameters, namely π11 = λ1 ω1 + λ2 ω2 (36)

π21 = π12 = λ2 ω1 π22 = 0

and of course ϑa = 0. Thus, the system is involution and the general solution, in the analytic category, depends on two functions of one variable. The symbol matrix σξ at a covector ξ = ξ1 ω1 + ξ2 ω2 is the 2 × 2 matrix ξ2 ξ1 . 0 ξ2 It follows that σξ has rank 2 except when ξ2 = 0, in which case, it has rank 1. It follows that the characteristic variety at each point of M is of the form Ξx = [ω1 ] ∈ P(Jx /Ix ). Of course, the characteristic sheaf will count this point with multiplicity 2. An example of this type of system is the pair of equations (37)

zy − wx = wy − z = 0.

Note that if we solve the first equation by introducing a potential function u so that z = ux and w = uy , then the remaining equation becomes the familiar parabolic equation ux = uyy . Thus, we shall say that systems which fall into this subcase are parabolic. Note that the system I has no Cauchy characteristics. However, the integrals of I are still foliated by the “characteristic curves” ω1 = 0. Moreover, on any integral of (I, Ω), we have π21 ∧ ω1 = 0, so π21 = λω1 for some smooth function λ. With this in mind, we define the Pfaffian system M on M to be the system spanned by the 1-forms ϑ1 , ϑ2, ω1 , π21 . It is easy to show that this span is well-defined independent of the choice of bases of J and I which put the structure equations of I in the form (35). Moreover, every integral of (I, Ω) is foliated by integral curves of M and these curves are clearly the characteristic curves. These are, of course, the characteristics in the classical sense of Monge (as well as in the modern sense).

236

VII. Examples

The case where the system M is of Frobenius type (so that the characteristics obey four “conservation laws”) is important because one can show that the system M is of Frobenius type if and only if the system I is locally diffeomorphic to the system given by the equations (38)

zy − wx = wy = 0.

This is one of those cases which Cartan described as “having characteristics depending on constants,” meaning that the maximal integrals of M depend only on constants (in this case, four constants). The fact that one can write down the general integral of such a system I by means of O.D.E. is no accident, but follows from a general procedure due to Darboux and Cartan whenever the system has characteristics depending on constants. Of course, one does not expect the general parabolic system to be solvable by O.D.E., and indeed one need only consider the system (37) whose solutions are equivalent to the solutions of the one-dimensional heat equation to see that the general solution cannot be written as an explicit function of two functions of one variable and a finite number of their derivatives. In this case, M fails to be a Frobenius system. Roughly speaking, the non-integrability of M corresponds to the original system of P.D.E. having parabolic “diffusion effects.” Subcase 2.4: The symbol relations are of type (26.4). In this case, by admissible basis change, we may assume that the symbol relations are of the form π21 ≡ C11 ω1 + C12ω2 (39) mod I. π12 ≡ C21 ω1 + C22ω2 Again, by adding appropriate multiples of ω1 and ω2 to the 1-forms πji , we may assume that the functions Cij are zero. Thus, the torsion vanishes and the structure equations now take the form dϑ1 ≡ −π11 ∧ ω1 (40) mod I. dϑ2 ≡ −π22 ∧ ω2 Inspection shows that s1 = 2 and s2 = 0. Moreover, the integral elements at a point of M clearly depend on 2 parameters, namely ϑa = 0 and πii = λi ωi and π21 = π12 = 0. Thus, the system is involutive and, in the analytic category, the general integral of (I, Ω) depends on two functions of one variable. The symbol matrix σξ at a non-zero covector ξ = ξ1 ω1 + ξ2 ω2 is the 2 × 2 matrix

ξ2 0

0 ξ1

.

It follows that the characteristic variety Ξx at each point x of M is the pair of points {[ω1 ], [ω2]} ∈ P(Jx /Ix ). A simple example of this type of system is the pair of equations (41)

zy = wx = 0.


237

The general solution is, of course, z = f(x) and w = g(y) where f and g are arbitrary functions of one variable. Note that this system is hyperbolic in the classical sense. In fact, it is easy to see that the general pair of equations whose symbol relations can be taken to be of the form (26.4) is hyperbolic in the classical sense. Thus, we shall call the equations which fit into this subcase hyperbolic. Again, there are no Cauchy characteristics. Nevertheless, the two characteristic foliations of an integral given by ω1 = 0 and ω2 = 0 define important geometric features of the integral manifolds. We shall not enter into a discussion of the classification of hyperbolic systems up to diffeomorphism. This would require a ´ Cartan which is too lengthy to enter discussion of the equivalence problem of Elie into here. However, we can indicate some basic invariants of the system I which can be used to determine whether a given system is equivalent to the “flat” system (41). Suppose, for the sake of convenience, that M is oriented. Then any local basis {ϑ1 , ϑ2} of I which satisfies the structure equations (40) and the condition ϑ1 ∧ dϑ1 ∧ ϑ2 ∧ dϑ2 > 0 is unique up to a change of basis of the form ϑã = λa ϑa where λ1 , λ2 = 0, as is easily verified. Using this, we can define two rank 3 Pfaffian systems M1 and M2 on M as follows. Write

(42.1)

dϑ1 dϑ2 dπ11 dω1

⎫ ≡0 ⎪ 2 2 ⎬ ≡ −π2 ∧ ω ≡ a1 π22 ∧ ω2 ⎪ ⎭ ≡ a2 π22 ∧ ω2

mod {ϑ1 , ϑ2 , ω1 , π11 }

(42.2)

dϑ2 dϑ1 dπ22 dω2

⎫ ≡0 ⎪ ⎬ ≡ −π11 ∧ ω1 ≡ a3 π11 ∧ ω1 ⎪ ⎭ ≡ a4 π11 ∧ ω1

mod {ϑ1 , ϑ2 , ω2 , π22 }

where the functions ai are some smooth functions locally defined on M . Replacing π11 , ω1 , π22 , and ω2 respectively by π11 + a1 ϑ2 , ω1 + a2 ϑ2 , π22 + a3 ϑ1 , and ω2 + a4 ϑ1 , we may assume that all of the ai are zero. We now have, in addition to (40), the refined structure equations (43.1)

dϑ1 ≡ dω1 ≡ dπ11 ≡ 0 mod {ϑ1 , ϑ2 , ω1 , π11 }

(43.2)

dϑ2 ≡ dω2 ≡ dπ22 ≡ 0 mod {ϑ1 , ϑ2, ω2 , π22 }.

It is now an elementary matter to see that the two Pfaffian systems (44.1)

M1 = span{ϑ1 , ω1 , π11 }

(44.2)

M2 = span{ϑ2 , ω2 , π22 }

are well-defined globally on M , that is, they depend only on the (hyperbolic) Pfaffian system I and the orientation of M (reversing the orientation of M switches the two systems). On each two dimensional integral N 2 of (I, Ω), each of the systems Ma restrict to have rank 1. Thus, each such N is foliated by integral curves of

238

VII. Examples

Ma . This pair of foliations is exactly the pair of foliations by characteristic curves in the classical sense. As in the parabolic case, we say that the characteristics of (I, Ω) depend on constants if both of the Ma are Frobenius systems. In this case, the integrals of (I, Ω) can locally be written down explicitly using only O.D.E. as follows: Supposing that each of the Ma is completely integrable, we know that M can be covered by open sets of the form U = Y 1 × Y 2 where each Y a is a three dimensional ball and the leaves of the Frobenius system Ma are given by fixing a point in the Y a -factor. It easily follows that ϑa is a non-zero multiple of a contact form, say ϑã, well defined on Y a . (The explicit construction of the factors Y a and the forms ϑã requires O.D.E. techniques.) It is now immediate that the integrals of (I, Ω) which lie in U are simply products of the form P 1 × P 2 where each P a is a contact curve in the contact manifold (Y a , ϑã). Clearly, such a system is equivalent to the “uncoupled” system (41). Actually, for the reduction of the initial value problem for integrals of (I, Ω) to an O.D.E. problem, it suffices that at least one of the systems Ma be Frobenius. Of course, for the general hyperbolic system, we do not expect either of the systems Ma to be Frobenius. In Darboux [1870], a far-reaching generalization of the above construction is presented which takes into account any possible “higher” conservation laws for characteristics.

Subcase 2.5: The symbol relations are of the form (26.5). In this case, we may, by admissible basis change, assume that the symbol relations are of the form π11 − π22 ≡ C11 ω1 + C12 ω2 (45) mod I. π21 + π12 ≡ C21 ω1 + C22 ω2 Again, by adding appropriate multiples of ω1 and ω2 to the 1-forms πji , we may assume that the functions Cij are zero. Thus, the torsion vanishes and the structure equations now take the form dϑ1 ≡ −π11 ∧ ω1 + π12 ∧ ω2 (46) mod I. dϑ2 ≡ −π12 ∧ ω1 − π11 ∧ ω2 Inspection shows that s1 = 2 and s2 = 0. Moreover, the integral elements of (I, Ω) at a point of M depend on 2 parameters, namely ϑ1 = 0 and (47)

π11 = π22 = λ1 ω1 + λ2 ω2 π21 = −π12 = λ2 ω1 − λ1 ω2

It follows that the system (I, Ω) is involution. In the analytic category, the general integral depends on two functions of one variable. The symbol matrix σξ at a non-zero covector ξ = ξ1 ω1 + ξ2 ω2 is the 2 × 2 matrix

ξ1 ξ2

−ξ2 ξ1

.


239

It follows that the characteristic variety Ξx at each point x of M is empty. On the other hand, the complex characteristic variety is not empty. In fact,√since det(σξ ) = 1 −1 ω2 ]. Again, (ξ1 )2 + (ξ2 )2 , it follows that ΞC x consists of the two points [ω ± this is in agreement with the predictions of the general theory for the dimension and degree of ΞC x. A simple example of a system whose symbol relations are of this type are the Cauchy–Riemann equations: (48)

zx − wy = zy + wx = 0.

In fact, it is easy to see that a pair of equations is elliptic in the classical sense if and only if its symbol relations are of the type given by the subspace (26.5). Thus, we shall refer to this type as elliptic. The “complexity” of the characteristic variety suggests that we study the system in a complex basis. Let us temporarily use the notation √ ϑ = ϑ1 + −1 ϑ2 √ (49) π = π11 + −1 π12 √ ω = ω1 + −1 ω2 . Then the structure equations (46) become (50)

dϑ ≡ −π ∧ ω dϑ¯ ≡ −¯ π∧ω ¯

mod I.

In fact, if we choose an orientation for M , then these equations plus the condition √ ¯ ϑ¯ > 0 uniquely specify ϑ ∈ I C up to a complex multiple. Moreover, −1 ϑ∧dϑ∧ ϑ∧d we have ⎫ dϑ ≡ 0 ⎪ ⎬ dϑ¯ ≡ −¯ π∧ω ¯ ¯ ω, π}. (51) mod {ϑ, ϑ, dπ ≡ a1 π¯ ∧ ω ¯⎭ ⎪ dω ≡ a2 π¯ ∧ ω ¯ ¯ we obtain, in addition to (49), the refined Replacing ω by ω +a2 ϑ¯ and π by π +a1 ϑ, structure equations (52)

dϑ ≡ dω ≡ dπ ≡ 0

¯ ω, π}. mod {ϑ, ϑ,

(Note the analogy with the hyperbolic case.) It is now an elementary matter to see that the complex Pfaffian system M spanned by the 1-forms {ϑ, π, ω} is well defined on M , depending only on the elliptic Pfaffian system I and the given orientation. ¯ Note (If one reverses the orientation, then the system M will be replaced by M.) that M defines a unique almost complex structure on M for which the system M is the space of forms of type (1, 0). By construction, the integrals of (I, Ω) are, in the terminology of Gromov [1985], “pseudo-holomorphic curves” for the given almost complex structure. In fact, the integrals of (I, Ω) are precisely the Mpseudo-holomorphic curves in M which are also integrals of the complex Pfaffian form ϑ (well-defined up to a complex multiple).

240

VII. Examples

The integrability of the almost complex structure M is the elliptic analogue of the notion of “characteristics depending on constants” in the hyperbolic case. By an application of the Newlander–Nirenberg theorem, it can be shown that the necessary and sufficient condition that the system I be locally equivalent to the one derived from the Cauchy–Riemann equations is that the system M be Frobenius (in the complex sense). This is a geometric indication of the special place the Cauchy– Riemann equations occupy in the study of elliptic systems for two functions of two variables. Case 3: r = 3. In the case where there are 3 symbol relations, we may choose bases of I and J as in (8)–(10) and the resulting 1-forms πia will satisfy 3 linear relations modulo J. It follows that, modulo J, we may assume that all of the 1-forms πia are multiples of a single 1-form π. Thus, we may write (53)

πia ≡ Raiπ

mod J.

The 2 × 2 matrix R may be changed by pre- and post-multiplication by invertible 2 × 2 matrices when we change to another adapted basis of I and J. Thus, the only invariant of R is its rank, which must be either 1 or 2. For the sake of simplicity, we shall assume that the rank of R is constant on M . This allows us to divide Case 3 into two subcases.

Subcase 3.1: The rank of R is 1. In this case, we may choose our bases of I and J so that the symbol relations become (54)

π21 ≡ π12 ≡ π22 ≡ 0 mod J.

(Note that we are reducing modulo J, not I, at this point). It follows that the structure equations for I can be written in the form (55)

dϑ1 ≡ −π11 ∧ ω1 + C 1 ω1 ∧ ω2 dϑ2 ≡ C 2 ω1 ∧ ω2

mod I

where C 1 and C 2 are smooth functions on M . Replacing π11 by π11 + C 1 ω2 , we see that we may assume that C 1 = 0. On the other hand, we clearly cannot get rid of the term involving C 2 , which represents the unabsorbable torsion. In fact, it is clear that there are no integrals of (I, Ω) outside of the locus where C 2 = 0. If C 2 = 0, then in the generic case, C 2 will vanish along a hypersurface in H in M . This locus may be regarded as a set of 4 equations for the two unknowns functions. We will return to this case below. We now pass on to the case where C 2 vanishes identically. In this case, the structure equations (55) simplify to (56)

dϑ1 ≡ −π11 ∧ ω1 dϑ2 ≡ 0

mod I.


241

Inspection shows that s1 = 1 and s2 = 0. Moreover, the integral elements at a general point of M depend on 1 parameter, namely ϑa = 0 and (57)

π11 = λω1 .

It follows that the system (I, Ω) is involution. In the real analytic case, the general integral depends on one function of one variable. The symbol matrix σξ at a non-zero covector ξ = ξ1 ω1 + ξ2 ω2 has the form ⎛

ξ2 ⎝0 0

⎞ 0 ξ1 ⎠ ξ2

Consequently, σξ is injective if and only if ξ2 = 0. Thus, at each point of M , Ξx consists of the point [ω1 ]. As a point of interest, although we shall not prove it here, we remark that there are locally only two systems of this kind up to diffeomorphisms which preserve the Pfaffian system I. The first is described by the equations (58)

zy = wx = wy = 0

and its general solution is given by z = f(x) and w = c where f is an arbitrary function of x and c is a constant. The second is described by (59)

zy = wx − z = wy = 0

with general solution z = f (x) and w = f(x) where f is an arbitrary function of x. The difference in the two systems is that, for the former, the first derived system I (1) is a Frobenius system while for the latter, I (1) is not a Frobenius system.

Subcase 3.2: The rank of R is 2. This is the generic case for systems of 3 equations for two functions of two variables. In this case we may choose bases of I and J so that the matrix R becomes the identity matrix. Thus, the symbol relations take the form (56)

π11 − π22 ≡ π21 ≡ π12 ≡ 0

mod J.

Thus, writing τ for π11 , the structure equations may be written in the form (61)

dϑ1 ≡ −τ ∧ ω1 + C 1 ω1 ∧ ω2 dϑ2 ≡ −τ ∧ ω2 + C 2 ω1 ∧ ω2

mod I.

Replacing τ by τ + C 1 ω2 − C 2 ω1 , we see that we may assume that the functions C a vanish identically. Thus, the torsion of this system is identically zero. The structure equations now simplify to dϑ1 ≡ −τ ∧ ω1 (62) mod I. dϑ2 ≡ −τ ∧ ω2

242

VII. Examples

Inspection shows that s1 = 1 and s2 = 0. However, there is a unique integral element of (I, Ω) at each point of M , given by ϑa = τ = 0. Thus, the system is not involutive. The symbol matrix σξ at a non-zero covector ξ is the 3 × 2 matrix ⎛ ⎞ ξ1 ξ2 ⎝ ξ2 0 ⎠ . 0 ξ1 Consequently, σξ is always injective. Thus, the complex characteristic variety is empty. Since there is a unique integral element at each point of M , the prolongation of (I, Ω) is particularly easy to compute. We may identify the space of integral elements of (I, Ω) with M itself and the differential system I (1) is just the Pfaffian system I+ of rank 3 which is generated by the 1-forms ϑ1 , ϑ2 , and τ . Since I+ is a rank 3 Pfaffian system on a manifold of dimension 5, we know that I+ is involutive if and only if it is a Frobenius system. By the structure equations (62), we already know that dϑa ≡ 0 mod I+ . It is clear that there exists a function C on M so that dτ ≡ Cω1 ∧ ω2 mod I+ . This function C vanishes if and only if the system I+ is differentially closed. In this case, i.e., C ≡ 0, the integrals of the system (I, Ω) clearly depend on 3 constants. In the generic case, however, not only will C not be identically zero, the locus where C = 0 will define a hypersurface in M which represents yet a fourth first order equation which must be adjoined to the given three. We will consider this case below. As an example of this type of system of P.D.E., let us consider the problem introduced at the beginning of this section of determining the isometries between two metrics on regions of the plane (see the discussion in the paragraph containing equations (16) through (18). As equation (18) shows, this system falls into Subcase 3.2. In order to investigate the closure of the related system I+ , we shall explicitly parametrize the group O(2). Recall that O(2) has two components, each of which is diffeomorphic to a circle. Thus, the general element of O(2) can be written in the form cos t sin t (63) g= . ∓ sin t ± cos t It then follows from equation (17) that the matrix π is skew-symmetric with upper right-hand entry π21 = dt ± ϕ12 − ψ21 . (The ±-sign is to be taken in agreement with the same sign in (61).) If we set τ = dt ± ϕ12 − ψ21 , then the differential system I+ is generated by ϑ1 , ϑ2 , and τ . We have the well-known formulae dϕ12 = K(z, w)η 1 ∧ η 2 dψ21 = k(x, y)ω1 ∧ ω2 , where K is the Gauss curvature of the metric h2 and k is the Gauss curvature of the metric h1 . Using the easily computed identity η 1 ∧ η 2 ≡ ±ω1 ∧ ω2 mod I, we obtain dτ ≡ (K − k)ω1 ∧ ω2 mod I+ .

§2. Finiteness of the Web Rank

243

Thus, the function C = K − k plays the role of the torsion of I+ . The function C can vanish identically only if each of the functions K and k are equal to the same constant. Thus, we recover the well-known fact that there exists a three-parameter family of local isometries between two metrics in the plane if and only if they each have the same constant Gauss curvature. More generally, since any integral manifold of (I, Ω) must lie in the locus C = 0, we see that any isometry between two such metrics must preserve their Gauss curvatures. It would be possible to pursue the study of this system further and arrive at a complete answer to the problem of determining when two metrics on the plane are locally isometric, but since our interest in this problem is only in the fact that it provides an example of a system in Subcase 3.2, we shall not go further into its analysis. Note, however, that this problem is not generic in the above sense because the locus C = 0 actually represents an equation of order zero relating the unknown functions z and w to the independent variables x and y.

Case 4: r = 4. In this case there are 4 independent relations of the form (11). This means that we may write these relations in the form (64)

πia ≡

a j Cij ω

mod I.

It follows that the structure equations of I are of the form (65)

a a dϑa ≡ (C12 − C21 )ω1 ∧ ω2

mod I.

Thus, it follows that the torsion of the system is represented by the two functions a a C a = C12 − C21 . These functions vanish identically if and only if the system I is a Frobenius system. In particular, I is Frobenius if and only if it is involutive. At the other extreme, in the generic case, the equations C 1 = C 2 = 0 implicitly define the functions z and w as functions, say f and g, of x and y. If these functions f and g do not solve the given equations, then there is no solution. æ §2. Finiteness of the Web Rank. The theory of webs had its origin in algebraic geometry and the theory of abelian integrals. For a more detailed introduction to the theory than we provide here, the reader should consult Chern and Griffiths [1978]. For our purposes, the following description will suffice. Let N n be a smooth manifold of dimension n. A d-web of codimension r on N is a d-tuple of foliations W = (F1 , F2 , . . . , Fd ) on N , each of codimension r, which satisfies the condition that the foliations Fa are pairwise transverse. (Actually, in algebraic geometry, this transversality condition is only supposed to hold on a dense open set in N .) Associated to each foliation Fa is the Pfaffian system Ia of rank r consisting of those 1-forms which vanish on the leaves of Fa . Of course, since Fa is a foliation, the Pfaffian system Ia is a Frobenius system. For each non-negative integer p which is less than or equal to r, we let

244

VII. Examples

Ωp (Ia ) denote the space of p-forms on N which may be expressed locally as sums of products of elements of Ia . Thus, if x1a , x2a, . . . , xra is a set of local functions whose differentials span Ia on the domain of their definition, then every Φ ∈ Ωp (Ia ) can be expressed uniquely on this domain in the form (66) Φ= fJ dxJa |J|=p

for some functions fJ . Given a d-web of codimension r, W, an abelian p-equation for W is a d-tuple (Φ1 , Φ2 , . . . , Φd ) where each Φa ∈ Ωp (Ia ) is a closed form and the d-tuple satisfies (67)

Φ1 + Φ2 + · · · + Φd = 0.

The space of all such abelian p-equations for W obviously forms a vector space which we shall denote by Ap (W). For the webs which occur in algebraic geometry, the dimension of this vector space is finite for global reasons. This dimension is usually called the abelian p-rank of W, and is an important invariant of W. It turns out that the finiteness of the abelian p-rank of W is a consequence of our general theorems about the characteristic variety of a differential system. In particular, the finiteness result we shall prove does not depend on any global conditions. For simplicity, we shall restrict ourselves to a discussion of the 1-rank. However, we will generalize the definition of web somewhat. Let us define a dpseudoweb W on N to be a d-tuple (I1 , I2 , . . . , Id ) of (non-singular) Pfaffian systems on N which are everywhere transverse: (Ia )x ∩ (Ib )x = 0 for all x ∈ N and all a = b. Note that we do not assume that the Pfaffian systems Ia are Frobenius and we do not assume that they have the same rank. Just as above, we define the vector space of abelian 1-equations A1 (N, W) to be the space of d-tuples (η1 , η2 , . . . , ηd) of closed 1-forms on N with ηa ∈ Ia and satisfying (68)

η1 + η2 + · · · + ηd = 0.

We then have the following proposition. Proposition 2.1. If N is connected, the dimension of A1 (N, W) is finite. Proof. Let V be a real vector space of dimension n and let U be an open neighborhood in N on which there exists a V -valued coordinate system y : U → V . Let ra be the rank of the Pfaffian system Ia and let Sa be a real vector space of dimension ra . Let Aa (y) : V → Sa be a surjective linear map (depending on y) so that ωa = Aa (y)dy is an Sa -valued 1-form satisfying the condition that the components of ωa span Ia |U . Let X = Rd × (S1 )∗ × (S2 )∗ × . . . × (Sd )∗ × U and let z1 , z2 , . . . , zd be coordinates on the Rd -factor while pa : X → (Sa )∗ is the projection onto the (Sa )∗ factor. Let M ⊂ X be the sub-locus defined by the n equations (69) pa Aa = 0. a

(Note that the left hand side of (4) is a function on X with values in V ∗ .) Let I denote the Pfaffian system on M which is generated by the d 1-forms ϑa where (70)

ϑa = dza − pa ωa = dza − (pa Aa )dy.

§3. Orthogonal Coordinates

245

We let I denote the differential ideal generated by I and we let Ω be the independence condition Ω = dy1 ∧ dy2 ∧ · · · ∧ dyn . Then it is clear that (I, Ω) is in linear form. An integral of (I, Ω) is a submanifold of M of the form za = fa (y) and pa = ga (y) on which the forms ϑa vanish. On such an integral, the forms ηa = ga (y)Aa (y)dy clearly satisfy (η1 , η2 , . . . , ηd) ∈ A(U, W). Conversely, given (η1 , η2 , . . . , ηd) ∈ A(U, W), there exist unique functions ga (y) with values in (Sa )∗ so that ηa = ga(y)Aa (y)dy. Also, since each of the forms ηa is closed, there exist functions fa , unique up to additive constants, so that ηa = dfa . It follows that the map given by (fa , ga ) → (ga (y)Aa (y)dy), from the integrals of (I, Ω) with domain N to A(U, W) is a surjective vector space mapping whose kernel consists of the vector space of dimension d given by setting all of the fa equal to constants and the ga equal to zero. Thus, in order to show that A(U, W) has finite dimension, it suffices to show that the space of integrals of (I, Ω) with domain N is a finite dimensional space. In turn, in order to show this, by Theorem 3.12 of Chapter V, it will suffice to show that the associated complex characteristic variety is empty. To prove that the complex characteristic variety is empty, we examine the tableau of (I, Ω). By the above description, if we let W = Rd , then the tableau at a point m = (y, z, p) ∈ M is the space & . & (71) Am = (ta Aa (y)) ∈ W ⊗ V ∗ & ta Aa = 0 a

(here the variable ta runs over the vector space (Sa )∗ . If ξ ∈ V ∗ were a non-zero covector such that [ξ] were characteristic, then there would be a non-zero w ∈ W so that w ⊗ ξ ∈ Am . Writing w = (wa ), this becomes wa ξ = ta Aa for some set of ta ∈ (Sa )∗ where the sum w1 + · · · + wd = 0. By this latter condition, it follows that at least two of the wa are non-zero. However, if wa = 0, then we clearly have ξ ∈ A∗a ((Sa )∗ ) ⊂ V ∗ . Since the assumption of transversality of the Pfaffian systems Ia clearly implies that A∗a ((Sa )∗ ) ∩ A∗b ((Sb )∗ ) = 0 for all a = b, we are lead to a contradiction. Thus, there cannot be any (complex) characteristic covectors. Thus, ΞC m is empty. We remark that the problem of determining good upper bounds for the dimension of A1 (N, W) for a general d-web of codimension r is rather subtle, being connected with Castelnuovo’s bound and the theory of special divisors in algebraic geometry. See Chern and Griffiths [1978]. §3. Orthogonal Coordinates. This example is a continuation of Example 3.2 of Chapter III and of Example 1.3 of Chapter V. We will follow the notation developed there. Recall that we are given a Riemannian metric g on a manifold N of dimension n and we wish to know when there exist local coordinates (called orthogonal coordinates) x1 , x2 , . . . , xn on N so that the metric takes the diagonal form (72)

g = g11 (dx1 )2 + g22 (dx2 )2 + · · · + gnn (dxn )2 .

As in Example 3.2 of Chapter III, we let F → N denote the orthonormal frame bundle of the metric g and we let ωi , ωij = −ωji denote the canonical 1-forms on

246

VII. Examples

F . These forms satisfy the structure equations dωi = −

ωij ∧ ωj

j

(73)

dωij = −

ωik ∧ ωkj +

k

1 2

Rijklωk ∧ ωl .

k,l

We saw, in the aforementioned example, that our problem was equivalent to finding integrals of the differential system I generated by the 3-forms Φi = ωi ∧ dωi subject to the independence condition Ω = ω1 ∧ ω2 ∧ · · · ∧ ωn . Moreover, we saw that the space of integral elements of (I, Ω) at a point of F was naturally an affine space of dimension n2 − n. Namely, if {pij }i =j is any collection of n2 − n real numbers, and f ∈ F is fixed, then the n)plane in Tf F annihilated by the 12 (n2 − n) 1-forms ϑij = ωij + pij ωi − pji ωj is an integral element of (I, Ω) and conversely, every integral element of (I, Ω) based at f is of this form for some unique collection of real numbers {pij }i =j . It follows that the first prolongation of (I, Ω) may be described in the following simple manner: Let F (1) = F × Rn(n−1) and let {pij }i =j be a set of (linear) coordinates on the second factor. Then I (1) is the differential system generated by the 12 (n2 − n) 1-forms ϑij = ωij + pij ωi − pji ωj . We may now compute the exterior derivatives of these 1-forms as follows: First, set Ωij = dωij + ωik ∧ ωkj . Then, we have dϑij = dpij ∧ ωi − dpji ∧ ωj + pij ∧ dωi − pji ∧ dωj + Ωij − ωik ∧ ωkj k

(74)

≡ dpij ∧ ωi − dpji ∧ ωj + + Ωij +

[pij pik ωi ∧ ωk − pji pjk ωj ∧ ωk ]

k

(pik ωi − pki ωk ) ∧ (pjk ωj − pkj ωk ) mod I (1)

k

≡π ˜ij ∧ ωi − π ˜ji ∧ ωj + Ωij mod I (1) where we have set, for all i = j, (75)

π ˜ij = dpij −

[(pij − pkj )pik ωk − 12 pik pjk ωj ]. k

It follows from (74) that, on an integral manifold of I (1), the 4-form Ωij ∧ ωi ∧ ωj must vanish for all i = j. Recalling from the Cartan structure equations that (76)

Ωij =

1 2

Rijklωk ∧ ωl ,

k,l

we see that, at any coframe f ∈ F which is part of an integrable orthonormal coframe, we must have Rijkl(f) = 0 whenever all of i, j, k, and l are distinct.

§3. Orthogonal Coordinates

247

Of course, if n = 3, then this last condition is trivially fulfilled at all coframes f. Moreover, we have already seen in §3 of Chapter III that when n = 3 the system (I, Ω) is involutive on F . By the prolongation theorem of §2 of Chapter VI, it follows that (I (1), Ω) is involutive in this dimension. Moreover, as computed in Example 1.3 of Chapter V, the characteristic variety at each integral element consists of 3 lines in general position. On the other hand, if n ≥ 4, then the structure equations of the Pfaffian system I (1) on F (1) are written in the form ˜ij ∧ ωi − π ˜ji ∧ ωj + 12 Rijklωk ∧ ωl mod I (1) . (77) dϑij ≡ π k,l

This is clearly a system in linear form. By the above remark, the torsion of this system does not vanish at any point (f, p) ∈ F (1) where Rijkl(f) = 0 for some quadruple of distinct indices (i, j, k, l). Rather than try for an exhaustive treatment, let us just consider the case where g is a metric on M with the property that Rijkl vanishes identically as a function on F whenever i, j, k, and l are distinct. Because this is a linear, constant coefficient system of equations on the Riemann curvature tensor which must hold in all orthonormal coframes, it follows that this condition is equivalent to assuming the vanishing of a certain number of the irreducible components of the Riemann curvature tensor under the action of the orthogonal group. Now it is well-known (see Besse [1986]) that for n ≥ 4, Kn , the space of Riemann curvature tensors in dimension n, decomposes into 3 irreducible subspaces under the action of O(n). These subspaces correspond to the scalar curvature, the traceless Ricci curvature, and the Weyl curvature. This gives an O(n)-invariant decomposition of a general Riemann curvature tensor into the form (78)

Rijkl = R(δik δjl − δil δjk ) + (Sik δjl − Sil δjk + Sjl δik − Sjk δil ) + Wijkl ,

where R is a scalar, Sij = Sji and satisfies i Sii = 0, and Wijkl has the same symmetries as the Riemann curvature tensor but in addition satisfies the trace condition i Wijik = 0. It is clear that if W (the Weyl curvature) vanishes, then Rijkl vanishes whenever all of the indices i, j, k, and l are distinct. Moreover, it is not difficult to exhibit a Weyl curvature tensor which satisfies W1234 = 0. It follows that the necessary and sufficient condition that Rijkl vanish identically in all frames whenever i, j, k, l are distinct is that the Weyl component of the curvature be zero. Thus, we shall assume W ≡ 0 from now on. It is not difficult to show that when n ≥ 4 this condition is equivalent to the condition that the metric g be conformally flat (see Besse [1986]). Since we are assuming that the Weyl curvature is zero, we have the formula (79)

Rijkl = Hik δjl − Hil δjk + Hjl δik − Hjk δil

where, for simplicity, we have set Hij = Sij + 12 Rδij . This gives the following simple formula for the curvature form Ωij : (80)

Ωij =

k

[Hik ωk ∧ ωj − Hjk ωk ∧ ωi ].

248

VII. Examples

It follows that by writing πij for π ˜ij − for I (1) simplify to the equations (81)

k

Hjk ωk for i = j, the structure equations

dϑij ≡ πij ∧ ωi − πji ∧ ωj

mod I (1) .

Thus, the torsion of I (1) is zero whenever the Weyl tensor vanishes identically. We are now going to show that the system (I (1) , Ω) is involutive. We begin by calculating the space of integral elements of (I (1) , Ω). On an integral element of at a given point of F (1) , the 2-forms (I (1), Ω) πij ∧ ωi − πji ∧ ωj must vanish even though the 1-forms ωi must remain linearly independent. It follows, in particular, that πij ∧ ωi ∧ ωj must vanish also whenever i = j. Thus, on any integral element E, πij must be a linear combination of the two 1-forms ωi and ωj . Let us write (82)

πij = Aij ωi + Bij ωj

for this linear relation which holds on E. Substituting (82) into the 2-form πij ∧ ωi − πji ∧ ωj yields the 2-form (Bij + Bji )ωi ∧ ωj . Thus, it follows that the matrix B must be skew-symmetric. Conversely, if Aij and Bij = −Bji are any collection of 3 2 2 (n − n) numbers (remember that we always have i = j), then the relations (82) and the conditions ϑij = 0 clearly suffice to define an integral element of (I (1) , Ω) at every point of F (1). Thus, the space of integral elements of (I (1), Ω) at each point of F (1) is an affine space of dimension 32 (n2 − n). To complete the proof of involutivity, it will suffice to show that we have s1 = s2 = 12 (n2 − n) while sk = 0 for all k > 2. We will do this by explicitly computing the polar equations for a pair of vectors v and w lying in the integral element E. Let e1 , e2 , . . . , en be the basis of E which is dual to ω1 , ω2 , . . . , ωn . Let v = a1 e1 + · · · + an en and w = b1 e1 + · · · + bn en where ak , bk are (at the moment) an arbitrary set of 2n numbers. The polar equations of the zero-dimensional subspace E0 ⊂ E are clearly the 1-forms ϑij . Thus, we have s0 = 12 (n2 − n). The polar equations of the 1-dimensional subspace E1 ⊂ E which is spanned by v consists of the forms ϑij and forms αij which satisfy αij ≡ ai πij − aj πji mod ω. As long as none of the numbers ai are zero, the rank of the polar equations of E1 is clearly (n2 − n). It follows that we must have s1 = 12 (n2 − n). Now let E2 ⊂ E be the vector space spanned by v and w. (In order for this space to be of dimension 2, we must have ai bj − aj bi = 0 for at least one pair i = j). The polar equations of E2 are spanned by the forms ϑij , the forms αij , and 1-forms βij which satisfy βij ≡ bi πij − bj πji mod ω. It follows that if ai bj − aj bi = 0 for all pairs i = j, then the rank of the polar equations of E2 will be 3 · 12 · (n2 − n). Of course, this implies that s2 = 12 · (n2 − n) and that sk = 0 for all k > 2 (since 3 · 12 · (n2 − n) is the codimension of E itself). Thus, (I (1), Ω) is involutive. It is interesting to compute the characteristic variety of this differential system. Appealing to Theorem 3.2 of Chapter V, we see that since s2 = 12 · (n2 − n) and sk = 0 for all k > 2, it follows that the characteristic variety is a curve of degree 1 2 n−1 . From the above computation of the characters, it follows that 2 · (n −n) in RP a 2-plane E2 ⊂ E is regular if and only if it has a basis v and w as above satisfying the condition ai bj − aj bi = 0 for all pairs i = j. In other words, E2 is regular if and only if the 2-forms ωi ∧ ωj are non-zero for all i = j. It follows immediately that a covector ξ = ξ1 ω1 + · · · + ξn ωn is characteristic if and only if it is of the form

§4. Isometric Embedding

249

ξ = ξi ωi + ξj ωj for some pair (i, j). Thus, the characteristic variety consists of the 1 2 ∗ 2 · (n − n) lines joining the n points [ωi ] ∈ P(E ). In closing, we note that since every metric in dimensions greater than 3 for which the Weyl tensor vanishes is conformally flat, and since a conformal change of metric does not change the status of orthogonal coordinates, we see that in the case where the Weyl curvature vanishes, we may assume that the metric g is actually flat. In particular, we may assume that the metric is real analytic. Then the Cartan–Kähler theorem may be applied to show that the space of local orthogonal coordinates on flat space depends on 12 · (n2 − n) functions of two variables. Note also that, as the theory predicts, the characteristic variety restricted to any solution is integrable. In fact, if x1 , x2 , . . . , xn are local orthogonal coordinates, then the characteristic hypersurfaces are those on which some pair of functions xi , xj become functionally dependent. §4. Isometric Embedding. In §3 of Chapter III, we applied differential systems to prove the Cartan–Janet embedding theorem. This theorem asserts that if g is a real analytic metric on a manifold N of dimension n, then g can be locally induced by local embeddings into the Euclidean space En+r for any 1 r ≥ 2 · n · (n − 1). In this section, we develop the application of differential systems to the study of the isometric embedding problem in the “overdetermined” case r < 12 · n · (n − 1). Here, prolongation and the characteristic variety play an important role. We shall adopt a slightly different approach to isometric embedding than that in Chapter III. There, we chose specific framings of various manifolds in order to avoid the complications which would have been induced into the calculations by the variable framings. Here, in order to simplify calculation of the characteristic variety and other geometric quantities, we will employ a more invariant formulation. We begin by reviewing the structure equations of a Riemannian metric. We shall adopt the index ranges 1 ≤ i, j, k ≤ n n + 1 ≤ a, b, c ≤ n + r 1 ≤ A, B, C ≤ n + r. Let g be a Riemannian metric on a manifold N of dimension n. Let x : F → N be the orthonormal frame bundle on N which consists of (n + 1)-tuples f = (x; e1 , e2 , . . . , en ) where x ∈ N and e1 , e2 , . . . , en is an orthonormal basis of Tx N . Of course, the group O(n) acts on F on the right in the usual way and makes F into a principal right O(n)-bundle over N . The tautological 1-forms ωi on F are defined by setting ωi (v) = g(x∗ (v), ei ) for all v ∈ Tf F with f = (x; e1 , e2 , . . . , en ). The Levi–Civita connection forms on F are the unique 1-forms ωij = −ωji which satisfy the first structure equation of Cartan (83)

dωi = −

j

ωij ∧ ωj .

250

VII. Examples

These forms also satisfy the second structure equations of Cartan (84)

dωij = −

ωik ∧ ωkj +

k

1 2

Rijklωk ∧ ωl

k,l

where the functions Rijkl are well defined on F and satisfy the symmetries (85)

Rijkl = −Rjikl = −Rijlk Rijkl + Riklj + Riljk = 0.

Just as in §3 of Chapter III, we shall regard the functions Rijkl as the components of a function R which takes values in the vector space Kn ⊂ Λ2 (Rn ) ⊗ Λ2 (Rn ) defined by the symmetries (85). Of course, the group O(n) acts linearly on Kn in the obvious way and the map R : F → Kn is equivariant. Let En+r be given its standard metric, and let F (En+r ) denote the orthonormal frame bundle of En+r . We shall denote elements of F (En+r ) as f = (y; e1 , e2 , . . . , en+r ). Of course, the group O(n + r) acts on the right on F (En+r ) and makes it into a principal right fiber bundle over En+r . We shall denote the tautological forms on F (En+r ) by ηA and the associated Levi–Civita forms by ηAB = −ηBA . Since En+r is flat, these forms satisfy the structure equations dηA = −

ηAB ∧ ηB

B

(86) dηAC = −

ηAB ∧ ηBC .

B

Let M be the manifold of 1-jets of isometries from N to En+r . Thus, an element of M consists of a triple m = (x, y, l) where x ∈ N , y ∈ En+r , and l : Tx N → Ty En+r is a linear map which is an isometry onto its image. We can embed M into Gn (T (N × En+r )) by letting m = (x, y, l) correspond to the n-plane E = {(v, l(v)) | v ∈ Tx N } which lies in T(x,y) (N × En+r ). The canonical Pfaffian system on Gn (T (N × En+r )) then restricts to M to be the Pfaffian system I0 of rank n + r which we wish to study. We let J0 be the canonical independence bundle. It contains I0 and is of rank 2n + r. In fact, J0 is the bundle of semi-basic 1-forms for the projection M → N × En+r . The difficulty of working directly with the Pfaffian system I0 on M is that there is no canonical set of generators for I0 on M . To remedy this, consider the product manifold F × F(En+r ). There is a canonical projection q : F × F(En+r ) → M defined by letting q(f, f ) be the triple (x, y, l) where x ∈ N is the base point of f, y ∈ En+r is the base point of f , and l : Tx N → Ty En+r is the linear map which satisfies l(ei ) = ei . Let O(n) × O(r) be the subgroup of O(n + r) which preserves the n-plane spanned by the first n elements of a given orthonormal basis of En+r . The group O(n) × O(r) acts on F × F(En+r ) by the diagonal action in the O(n)factor and in the standard way in the O(r)-factor. It is clear that the orbits of O(n) × O(r) are the fibers of q. Thus, F × F(En+r ) has the structure of a principal right O(n) × O(r)-bundle over M . It is easy to see that the Pfaffian system I0 on M pulls back up to F ×F(En+r ) to be spanned by the 1-forms {ηi −ωi | 1 ≤ i ≤ n} and the 1-forms {ηa | n < a ≤ n+r}.


251

Let I0 denote the differential system generated by I0 on either M or F × F(En+r ). The structure equations above give the formulas ⎫ d(ηi − ωi ) ≡ − (ηij − ωij ) ∧ ωj ⎪ ⎪ ⎪ ⎬ j (87) mod I0 ⎪ ⎪ dηa ≡ − ηaj ∧ ωj ⎪ ⎭ j

which make clear the fact that, on F × F(En+r ), the orbits of O(n) × O(r) are the Cauchy leaves of I0 . Moreover, the independence condition on M pulled up to F × F (En+r ) can be represented by the n-form Ω = ω1 ∧ ω2 ∧ · · · ∧ ωn . To include the Cauchy characteristics, we shall define an augmented independence condition on F × F(En+r ) by letting Ω+ be the form of degree n+ = n + 12 n(n − 1) + 12 r(r − 1) obtained by wedging together the forms ωi , {ωij | i < j}, and the forms {ηab | a < b}. It is now easy to see that every n+ -dimensional integral element E+ of (I0 , Ω+ ) on F ×F(En+r ) pushes down to M to be an n-dimensional integral element E = q∗ (E+ ) of (I0 , Ω). Conversely, for every point (f, f ) ∈ F ×F(En+r ) in the fiber over m ∈ M , the inverse image of an integral element E ⊂ Tm M of (I0 , Ω) is an integral element E+ = q∗−1 (E) ⊂ T(f,f ) F × F(En+r ) of (I0 , Ω+ ). By the equations (87), the integral elements of (I0 , Ω+ ) based at a point (f, f ) ∈ F × F(En+r ) are parametrized by a collection of 12 (n2 + n)r numbers h = (haij ) = (haji) in the following way: For each such collection, the n+ -plane which is annihilated by the 1-forms of I0 together with the 1-forms ηij − ωij and the 1-forms ηai − haij ωj is an integral element of (I0 , Ω+ ) and, conversely, every integral element of (I0 , Ω+ ) is of this form. Thus, the space Vn (I0 , Ω+ ) of integral elements of (I0 , Ω+ ) is diffeomorphic to the product F ×F(En+r )×(Rr ⊗S 2 (Rn )). Accordingly, we shall denote the elements of Vn (I0 , Ω+ ) by triples (f, f, h). Now, the group O(n) × O(r) acts in the obvious way on the vector space Rr ⊗ 2 S (Rn ). If we let O(n)×O(r) act on F ×F(En+r )×(Rr ⊗S 2 (Rn )) by the consequent natural diagonal action, then the quotient will clearly be the space of integral elements of (I0 , Ω) on M . Thus, M (1) = (F × F(En+r ) × (Rr ⊗ S 2 (Rn )))/O(n) × O(r). Geometrically, we may represent the elements of M (1) as quadruples (x, y, l, H) where (x, y, l) ∈ M is as before and H : S 2 (Tx N ) → (l(Tx N ))⊥ is a linear map. Here, (l(Tx N ))⊥ is the r-dimensional vector space which is perpendicular to l(Tx N ) in Ty En+r . As to be expected, the elements of M (1) are identifiable with the 2-jets of smooth maps which are isometries up to first order. (1) For simplicity, we shall denote by I (instead of I0 ) the Pfaffian system which is generated on F × F(En+r ) × (Rr ⊗ S 2 (Rn )) by the 1-forms ϑi , ϑa, ϑij = −ϑji , and ϑai where ϑ i = η i − ωi ϑa = ηa (88)

ϑij = ηij − ωij ϑai = ηai − haij ωj . j

252

VII. Examples

Of course, I is also well-defined as a Pfaffian system on M (1) and moreover, the Cauchy leaves of I on F × F(En+r ) × (Rr ⊗ S 2 (Rn )) are the orbits of the group O(n) × O(r). We let I denote the differential system generated by I on either M (1) or F × F(En+r ) × (Rr ⊗ S 2 (Rn )). The structure equations of I are easily obtained in the form ⎫ dϑi ≡ dϑa ≡ 0 ⎪ ⎪ ⎪ ⎪ 1 dϑij ≡ 2 Tijklωk ∧ ωl ⎬ (89) mod I ⎪ ⎪ dϑai ≡ − πaij ∧ ωj ⎪ ⎪ ⎭ j

where the forms πaij = πaji are given by the formula (90)

πaij = dhaij −

[hakj ωki + haik ωkj ] +

k

hbij ηba.

b

and the functions Tijkl are given by the formulas (91)

Tijkl =

[haik hajl − hail hajk ] − Rijkl.

a

It is clear from (89) that there are no integral elements of (I, Ω+ ) at points of F × F(En+r ) × (Rr ⊗ S 2 (Rn )) where any of the functions Tijkl are non-zero (the functions Tijkl represent non-absorbable torsion). The usual prescription for continuing the prolongation process at this point is to restrict to the subspace of F × F(En+r ) × (Rr ⊗ S 2 (Rn )) where all of the functions Tijkl vanish. If we let Z ⊂ F × F(En+r ) × (Rr ⊗ S 2 (Rn )) denote the locus of common zeros of the collection of functions T = {Tijkl | 1 ≤ i, j, k, l ≤ n}, and let q(Z) ⊂ M (1) denote its image in M (1) , then q(Z) represents the 2-jets of mappings of N to En+r which induce an isometry to first order and which satisfy the Gauss equations Tijkl = 0. (Note that, from a given 2-jet of an immersion into Euclidean space, one can only compute the 1-jet of the induced metric. The Gauss equations demonstrate the remarkable fact that, even though one needs the the full 3-jet of the immersion to compute the full 2-jet of the metric, the Riemannian curvature tensor of the induced metric (which depends only on partial 2-jet information) can be computed using only the 2-jet of the immersion.) Unfortunately, Z will not, in general, be a smooth manifold. To avoid this difficulty, we shall restrict our attention to a more manageable subspace. Let Z ⊂ Z denote the subspace consisting of the ordinary zeros of the collection T . (Recall that if P is a smooth manifold and P is a collection of smooth functions on P , then a point p ∈ P is called an ordinary zero of the collection P if there exists an integer k, an open neighborhood U of p, and a set of functions f1 , f2 , . . . , fk ∈ P whose differentials are linearly independent on U which have the property that an element q ∈ U is a simultaneous zero of all of the functions in P if and only if it is is a zero of the functions f1 , f2 , . . . , fk . The integer k is, of course, the codimension of the zero locus of the functions in P at p.) Moreover, since we are only interested in integrals


253

of (I, Ω+ ), we may as well restrict to Z ∗ ⊂ X, the open subset of Z on which the form Ω+ is non-zero. Assuming that Z ∗ is non-empty (as it will be in many of the specific examples below), we can study the restriction of the differential system (I, Ω+ ) to Z ∗ . Note that since the collection T and the form Ω+ are invariant (up to sign) under the action of the group O(n) × O(r), it follows that Z and Z ∗ are invariant under this action and thus are unions of its orbits. Let Y ⊂ M (1) be the image of Z under the quotient mapping. Then Y is a smooth submanifold of M (1) . Moreover, if Y ∗ ⊂ Y is the open subset where Ω does not vanish, then Y ∗ is clearly the image of Z ∗ under the quotient map. In fact, Z ∗ is an O(n) × O(r)-bundle over Y ∗. In order to study the restriction of the Pfaffian system I to Z ∗ or Y ∗ , we shall need some information about the differentials of the functions in T . Let us set (92) τijkl = dTijkl − [Tmjkl ωmi + Timkl ωmj + Tijml ωmk + Tijkmωml ]. m

The structure equations (84) can be differentiated to yield the following formula for the derivative of Rijkl dRijkl = [Rmjklωmi + Rimkl ωmj + Rijml ωmk + Rijkmωml + Rijklmωm ]. m

where the functions Rijklm are uniquely specified by this formula and represent the components of ∇R, the covariant derivative of the Riemann curvature tensor. If we regard hij = (haij ) as an Rr -valued function on F × F(En+r ) × (Rr ⊗ S 2 (Rn )) and use the standard inner product on Rr , then the formula for Tijkl can be simplified to the formula Tijkl = hik · hjl − hil · hjk − Rijkl. It follows that if we let πij = (πaij ) denote the corresponding Rr -valued 1-form, then the formula for the differentials of the functions Tijkl can be written in the form (93) τijkl = hik · πjl − hil · πjk + πik · hjl − πil · hjk − Rijklmωm . m

On Z, we have dTijkl = τijkl. If C is the codimension of Z in F ×F(En+r )×(Rr ⊗ S (Rn )) at (f, f , h) ∈ Z, then exactly C of the forms τijkl are linearly independent at (f, f , h) and the vanishing of the C corresponding functions Tijkl suffice to define Z in a neighborhood of (f, f , h). Moreover, since Z ∗ is an open subset of Z, it follows that, if (f, f , h) ∈ Z ∗ , then these C 1-forms must be linearly independent from the 1forms ωi . In particular, at points of Z ∗ , the number of linearly independent 1-forms among the τijkl is the same as the number of linearly independent 1-forms among the τijkl mod {ω1 , . . . , ωn }. This latter rank clearly depends only on the “second fundamental form” h. This justifies the following terminology: We shall speak of Y ∗ as the space of 2-jets of isometric immersions with ordinary second fundamental form. The local isometric embeddings of N into En+r which correspond to integrals of I on Z ∗ will be referred to as ordinary isometric embeddings. We now proceed to investigate the ordinary isometric embeddings. Upon restriction to Z ∗ , the new relations on the forms in the coframing 2

ϑi , ϑa, ϑij , ϑai , πij , ωi , ωij , ηab

254

VII. Examples

will all be generated by setting the differentials of the Tijkl equal to zero. Since, on Z, we have dTijkl = τijkl, it follows that the structure equations of I on Z ∗ are ⎫ ⎪ ⎪ ⎪ ⎪ ⎬

dϑi ≡ dϑa ≡ 0 (94)

dϑij ≡ 0 dϑai ≡ −

j

⎪ ⎪ πaij ∧ ωj ⎪ ⎪ ⎭

mod I

where the relations on the forms πij are spanned by πij = πji and (95)

hik · πjl − hil · πjk + πik · hjl − πil · hjk ≡

Rijklmωm

mod I.

m

Of course, one does not expect this system to be involutive in general. For one thing, we have not yet made any assumption about r, the embedding codimension. However, we can already gain some useful information about the isometric embedding problem by examining the characteristic variety of the symbol relations (95). Assume that E is an integral element of (I, Ω) at a point (x, y, l, H) of Y ∗ . We want to compute the condition that a covector ξ ∈ E ∗ be characteristic. Let (f, f , h) ∈ Z ∗ be an element which lies over (x, y, l, H) and has the property that ξ(q∗ (v)) = λωn (v) for some non-zero real number λ and every v ∈ E+ where E+ = q∗−1 (E) is the associated integral element of (I, Ω+ ) on Z ∗ . Since O(n) acts transitively on the unit sphere in Rn , such an element (f, f , h) clearly exists. The annihilator of E+ is spanned by the 1-forms in I together with some Rr -valued 1-forms (96) ψij = ψji = πij − hijk ωk , k

where the vectors hijk = hjik = hikj in Rr satisfy the equations (97)

hik · hjlm − hil · hjkm + hjl · hikm − hjk · hilm = Rijklm(f).

The polar equations of the hyperplane ωn = 0 in E+ are spanned by the 1-forms in I together with the 1-forms Ψ = {ψij | i < n and i ≤ j}. It follows that this hyperplane is characteristic if and only if not all of the components of the 1-form ψnn can be obtained as linear combinations of the components of the forms in Ψ. Now, because Z ∗ is an open subset of the space of ordinary zeros of the functions T , aside from the symmetry ψij = ψji , the only relations among the forms ψij are (98)

hik · ψjl − hil · ψjk + hjl · ψik − hjk · ψil = 0.

All of the relations in (98) which involve ψnn in a non-trivial way can be obtained by letting i and k be strictly less than n and setting j and l equal to n. This gives rise to the relations (99)

hik · ψnn ≡ 0 mod Ψ


255

for all pairs i and k which are strictly less than n. This set of relations implies ψnn ≡ 0 mod Ψ if and only if the set of vectors {hij | i, j < n} spans the entire + vector space Rr . Thus, the hyperplane ωn = 0 in E is characteristic if and only if, when we regard h as a quadratic form h = i,j hij ωi ◦ ωj with values in Rr , there exists a non-zero vector w ∈ Rr so that w · h is a multiple of ωn . Translating this into a statement about the original integral element E of (I, Ω) based at (x, y, l, H) ∈ Y ∗ , we see that, after we make the natural identification E∼ = Tx N , a covector ξ ∈ E ∗ is characteristic if and only if there exists a non-zero vector w ∈ (l(Tx N ))⊥ so that w · H = λ ◦ ξ for some λ ∈ E ∗ . We record this as the following fundamental proposition. Proposition 4.1. If E is an integral element of (I, Ω) based at (x, y, l, H) in Y ∗ , then a covector ξ ∈ E ∗ is characteristic if and only if there exists a non-zero vector w ∈ (l(Tx N ))⊥ so that w · H = λ ◦ ξ for some λ ∈ E ∗ . Of course, Proposition 4.1 can be applied directly only to the cases where the system (I, Ω) has been shown to be involutive. However, as discussed in Chapter V, the characteristic variety can only decrease in size when one prolongs. Thus, Proposition 4.1 serves as an important means of deriving an “upper bound” for the characteristic variety. Proposition 4.1 also serves as motivation for the definition of a sort of characteristic variety associated to any “second fundamental form.” Thus, let W be any r-dimensional Euclidean vector space and let V be any n-dimensional real vector space. Given an element H ∈ W ⊗ S 2 (V ∗ ), which we may regard as a quadratic form on V with values in W , we define ΞH ⊂ PV ∗ by the condition that [ξ] ∈ ΞH if and only if there exists some non-zero w ∈ W so that w · H = λ ◦ ξ for some λ ∈ V ∗ (note that we do not require λ to be non-zero). By tensoring with C, we may define 2 ∗ the associated complex variety ΞC H . It will also be useful to let |H| ⊂ S (V ) denote the linear subspace spanned by the quadratic forms w · H as w ranges over all of W. As a first application of Proposition 4.1, we consider the “underdetermined” case when r > 12 n(n − 1). Proposition 4.2. If r > 12 n(n − 1), then ΞH = PV ∗ for all H ∈ W ⊗ S 2 (V ∗ ). Proof. Note first that if dim |H| < r, then there exists a vector w ∈ W so that w·H = 0 and hence every covector is characteristic. On the other hand, if dim |H| = r, then for every non-zero ξ ∈ V ∗ , the linear space (ξ) = {λ ◦ ξ | λ ∈ V ∗ } must have non-trivial intersection with |H| for dimension reasons. Thus, again, every covector in V ∗ must be characteristic. The “determined” case is also quite interesting: Proposition 4.3. If r = 12 n(n − 1), then for all H ∈ W ⊗ S 2 (V ∗ ), ΞC H is either all of P(V C )∗ or else is an algebraic hypersurface in P(V C )∗ of degree n. Moreover, if n > 2, then the real characteristic variety ΞH is never empty. Proof. Again, if dim |H| < r, then every covector is characteristic and there is nothing to prove. Thus, we may assume for the rest of the proof that dim |H| = r. Let h1 , h2 , . . . , hr be a basis of |H| ⊂ S 2 (V ∗ ). Let x1 , x2 , . . . , xn be a basis of i ∗ V ∗ and let ξ = where we regard the ξi as i ξi x be a general element of V variables. Since S 2 (V ∗ ) is of dimension r + n, let ∆ be a basis of Λr+n (S 2 (V ∗ )).

256

VII. Examples

Then there exists a homogeneous polynomial P (ξ1 , ξ2 , . . . , ξn ) of degree n (with real coefficients) so that (100)

h1 ∧ h2 ∧ · · · ∧ hr ∧ (ξ ◦ x1 ) ∧ (ξ ◦ x2 ) ∧ · · · ∧ (ξ ◦ xn ) = P (ξ)∆.

Clearly, [ξ] ∈ P(V C )∗ is characteristic if and only if P (ξ) = 0. The stated properties of ΞC H now follow immediately. To show that ΞH is non-empty whenever n > 2, let us set Q = S 2 (V ∗ )/|H|. Then Q a real vector space of dimension n. For any element b of S 2 (V ∗ ), let us let [[b]] ∈ Q denote its reduction mod |H|. Suppose now that ΞH were empty. Then for any two non-zero elements α, β ∈ V ∗ , we must have α ◦ β ∈ / |H| and thus [[α ◦ β]] = 0. It follows that if we define µ : V ∗ × V ∗ → Q by µ(α, β) = [[α ◦ β]], then µ is a symmetric multiplication without zero divisors. Thus, if we choose any non-zero e ∈ V ∗ , then we may define a V ∗ -valued product α ◦ β by letting α ◦ β be the unique element of V ∗ which satisfies µ(e, α · β) = [[α ◦ β]]. This product defines the structure of a commutative (though not necessarily associative) algebra on V ∗ . By construction, this algebra has a unit e and has no zero divisors. By the commutativity of V ∗ , it follows that we may choose a basis of V ∗ in which all of the maps mα : V ∗ → V ∗ given by mα (β) = α · β are simultaneously in (real) Jordan canonical form. From this, it immediately follows that these maps cannot all be invertible unless n ≤ 2. Note that in the case n = 2, we have r = 1. Thus |H| consists of the multiples of a single quadratic form on V , say b. If b is positive (or negative) definite, then b = ±ξ ◦ ξ¯ where ξ is a complex-valued 1-form on V . Thus, in this case, ΞH consists ¯ in P(V C )∗ . The case of isometric embedding of the two (non-real) points {[ξ], [ξ]} when n = 2 and r = 1 has been discussed rather thoroughly in Chapters IV and V. We shall not discuss it further here. It is interesting to remark that considerations from algebraic geometry allow us to compute the dimension and degree of ΞC H for “generic” H. The dimension is easy: Since we have already treated the other cases, we may assume that r < 12 n(n − 1). Then for generic H ∈ W ⊗ S 2 (V ∗ ), we will have dim |H| = r, and |H| will be a generic r-plane in S 2 (V ∗ ). The cone C ⊂ S 2 (V ∗ ) which consists of quadratic forms of rank 2 or less is a singular cone of dimension 2n − 1. For dimension reasons, if r + 2n − 1 ≤ 12 n(n + 1) = dim S 2 (V ∗ ), then for generic H, |H|C ∩ C C = {0}. It then follows from Proposition 4.1 that if r ≤ 12 (n − 1)(n − 2) then, for generic H, 1 ΞC H = ∅. If r = k + 1 + 2 (n − 1)(n − 2) where 0 ≤ k < n − 1, then the above general position argument shows that, again for generic H, dim ΞC H = k. is somewhat more difficult to compute. We shall not reproduce The degree of ΞC H the argument here, instead we refer the reader to Bryant, Griffiths and Yang [1983], where it is shown that, for generic H, deg ΞC H

=

2n − 2 − k n−1

.

Returning to the case r = 12 n(n−1), let us say that an element H ∈ W ⊗S 2 (V ∗ ) is non-degenerate if ΞH is not all of PV ∗ . Note that the condition of non-degeneracy depends only on the subspace |H| ⊂ S 2 (V ∗ ). In order to have non-degeneracy, |H| must have dimension r. Moreover, as (100) shows, the coefficients of P (ξ) are linear


257

in the Pl¨ ucker coordinates of the subspace |H| in Gr (S 2 (V ∗ )). It follows that the set of non-degenerate elements of W ⊗ S 2 (V ∗ ) is an open subset U in W ⊗ S 2 (V ∗ ). We can say more. Suppose that H ∈ W ⊗ S 2 (V ∗ ) is non-degenerate and that ξ ∈ V ∗ is non-characteristic for H. Let x1 , x2, . . . , xn be a basis of V ∗ with the n property that ξ = x . We may expand H in the form H = i,j hij xi ◦ xj where the hij = hji are vectors in W . It follows immediately that the hypothesis that ξ not be characteristic is equivalent to the condition that the r vectors {hij | i ≤ j < n} be linearly independent. This leads to the following important observation. Recall that K(V ) ⊂ Λ2 (V ∗ ) ⊗ Λ2 (V ∗ ), the space of Riemann curvature tensors on V is defined to be the kernel of the natural map from Λ2 (V ∗ ) ⊗ Λ2 (V ∗ ) to V ∗ ⊗ Λ3 (V ∗ ) obtained by “skew symmetrizing on the last three indices.” (When V is explicitly identified as Rn , we use the already established notation Kn instead of K(Rn ).) It is well known that K(V ) is actually a subspace of S 2 (Λ2 (V ∗ )) ⊂ Λ2 (V ∗ ) ⊗ Λ2 (V ∗ ) and has dimension n2 (n2 − 1)/12. There is a natural map γ : W ⊗ S 2 (V ∗ ) → K(V ) which is defined in indices by setting (101)

γ(H)ijkl = hik · hjl − hil · hjk

where we have expanded H in indices as H = i,j hij xi ◦ xj . It is important to note that, since no inner product on V is used in the definition of γ, this map is equivariant under the action of the group O(W ) × GL(V ). We shall return to this point later. Referring to Lemma 3.10 of Chapter III, we have the proposition Proposition 4.4. If r = 12 n(n − 1) and U ⊂ W ⊗ S 2 (V ∗ ) is the open set consisting of non-degenerate elements, then γ : U → K(V ) is a surjective submersion. This leads to the following strengthening of Theorem 3.11 of Chapter III. Proposition 4.5. If r = 12 n(n − 1) and Z ⊂ F × F(En+r ) × (Rr ⊗ S 2 (Rn )) is the set of zeros of the collection T = {Tijkl }, then ZU = Z ∩ (F × F(Cn+r ) × U) consists entirely of ordinary zeros of T . Moreover, ZU is an open subset of Z ∗ on which the differential system (I, Ω+) is involutive with Cartan characters sp = 1 ∗ 2 n(n − p)(n − p + 1) for p ≤ n on the open subset ZU /(O(n) × O(r)) ⊂ Y . Proof. The function T : F × F(En+r ) × (Rr ⊗ S 2 (Rn )) → Kn with components Tijkl can be written in the form T = γ − R where γ : Rr ⊗ S 2 (Rn ) → Kn is defined above and R : F → Kn is the Riemann curvature tensor. Since γ is a surjective submersion when restricted to U ⊂ Rr ⊗ S 2 (Rn ), it follows that T is a surjective submersion when restricted to F ×F(En+r )×U. Of course, this immediately implies that any zero of T which lies in F × F(En+r ) × U must be an ordinary zero of T . This establishes that ZU ⊂ Z. Moreover, since T is a surjective submersion when restricted to any fiber of the form {f} × {f } × U, it follows that the projection ZU → F × F(En+r ) is a surjective submersion. Thus, it follows that Ω+ does not vanish on ZU . In particular, ZU ⊂ Z ∗ . The proof of Theorem 3.11 can now easily be adapted to show that ZU is a open submanifold of the space of ordinary integral elements of the differential system with independence condition (Io,+ , Ω+ ) on F × F(En+r ), where Io,+ is the differential ideal generated by the 1-forms in Io together with the 1-forms ϑij = ηij − ωij . Moreover, that proof further shows that (Io,+ , Ω+ ) is involutive on F × F(En+r ) with Cartan characters s¯p = 12 n(n − p) for p ≤ n and s¯p = 0 for n < p ≤ n+ .

258

VII. Examples

Thus, the differential system (I, Ω+ ) on ZU is seen to be (an open subset of) the ordinary prolongation of the involutive system (Io,+ , Ω+ ). By Theorem 2.1 of Chapter VI, it follows that (I, Ω+ ) is involutive on ZU with Cartan characters sp = 12 n(n − p)(n − p + 1) for p ≤ n and sp = 0 for n < p ≤ n+ . The remainder of the proposition now follows upon quotienting by the Cauchy leaves. Combining Propositions 4.3–5, we arrive at the following result for isometric embedding in the determined case r = 12 n(n − 1). Theorem 4.6. The differential system for isometrically embedding a given metric on N n into the Euclidean space of dimension n + 12 n(n − 1) = 12 n(n + 1) in such a way that the second fundamental form is non-degenerate is an involutive differential system with independence condition (I, Ω) on the manifold of 2-jets of immersions of N n into En(n+1)/2 which are infinitessimal isometries, satisfy the Gauss equations, and induce non-degenerate second fundamental forms. Moreover, the characters of this system are sp = 12 n(n − p)(n − p + 1) for all p ≤ n. Given such a non-degenerate isometric immersion, u : N → En(n+1)/2, the comC plex characteristic variety ΞC x ⊂ P(Tx N ) is an algebraic hypersurface of degree n. Moreover, if n ≥ 3, then the real characteristic variety Ξx is non-empty for all x ∈ N. Proof. Omitted.

Note that, as a consequence of the non-emptyness of the real characteristic variety when n ≥ 3, it follows that the determined isometric embedding problem is never elliptic for n ≥ 3, a result first noted by Tanaka [1973]. However, this does not mean that the isometric embedding problem in the smooth category is out of the reach of analysis when n ≥ 3. In fact, in Bryant, Griffiths and Yang [1983], it is shown that a careful study of the characteristic variety when n = 3 can be coupled with the Nash–Moser Implicit Function Theorem to prove the existence of a smooth isometric embedding N 3 → E6 in some neighborhood of any point x ∈ N for which the Riemann curvature tensor is suitably non-degenerate. By applying more subtle results from the theory of P.D.E. of principal type, Yang and Goodman have been able to weaken this non-degeneracy assumption to the assumption that the Riemann curvature tensor be non-zero at x. At present, it is unknown whether there exists a local embedding on a neighborhood of a point where the Riemann curvature tensor vanishes. The analytical difficulties are similar to (but more complicated than) the difficulties one faces in trying to isometrically embed an abstract surface N 2 into E3 on a neighborhood of a point where the Gauss curvature vanishes. Also, Yang has some results on smooth solvability in the case n = 4. These results depend on the theory of P.D.E. of real principal type and show existence of an isometric embedding in a neighborhood of a point x ∈ N 4 where the metric satisfies some open condition on a finite jet of the metric at that point. For more details, see Goodman and Yang [1990]. In the cases n ≥ 5, the analysis is complicated by the fact that the characteristic variety is generally no longer smooth (Bryant, Griffiths and Yang [1983]). We now turn to the overdetermined cases r < 12 n(n − 1). In these cases, the Gauss equations represent non-trivial obstructions to isometric embedding. For example, when n = 2, isometric embeddings satisfying the condition r = 0 exist only when the Gauss curvature of the metric g vanishes identically. Since we have already dealt with this case in §1, we shall henceforth assume that n > 2.


259

A fundamental part of the problem is to study the mapping γ : W ⊗ S 2 (V ∗ ) → K(V ). Since the dimensions of S 2 (V ∗ ) and K(V ) are 12 n(n + 1) and n2 (n2 − 1)/12 respectively, it is clear that when n is large, the value of r = dim W must also be large if we are to have surjectivity of the map γ. Let us make a detailed study of the case n = 3. Since, K(V ) has dimension 6, it follows that K(V ) = S 2 (Λ2 (V ∗ )). Thus, K(V ) can be regarded as the space of quadratic forms on Λ2 (V ). The values of r which we will be interested in are r = 1 and r = 2. In the case r = 1, the space W ⊗ S 2 (V ∗ ) is identified with S 2 (V ∗ ), the space of quadratic forms on V . If h ∈ S 2 (V ∗ ), we may diagonalize h in the form h = λ1 (x1 )2 + λ2 (x2 )2 + λ3 (x3 )2 , where x1 , x2 , x3 is a basis of V ∗ and the λi are real numbers. It then follows that γ(h) = λ2 λ3 (x2 ∧ x3 )2 + λ3 λ1 (x3 ∧ x1 )2 + λ1 λ2 (x1 ∧ x2 )2 as an element of S 2 (Λ2 (V ∗ )) = K(V ). Note that, as a quadratic form, γ(h) has type (3, 0), (1, 2), (1, 0), (0, 1), or (0, 0). Conversely, any quadratic form on Λ2 (V ) whose type belongs to this list can be realized as γ(h) for some h ∈ S 2 (V ∗ ). In fact, the map γ can be understood quite easily in terms of matrices. If we fix a basis x1 , x2 , x3 of V ∗ , then we may identify elements h in S 2 (V ∗ ) with 3 × 3 symmetric matrices in the usual way. Using the corresponding basis x2 ∧x3 , x3 ∧x1 , x1 ∧ x2 of Λ2 (V ∗ ), we may also identify elements K in S 2 (Λ2 (V ∗ )) = K(V ) with 3 ×3 symmetric matrices. Under these identifications, it is easy to see that the map γ becomes identified with the map Adj, which associates to each 3 × 3 symmetric matrix its adjoint matrix (i.e., the matrix of signed 2×2 minors). Since the identities det(Adj(h)) = (det(h))2 Adj(Adj(h)) = det(h) · h hold for all 3 × 3 symmetric matrices, it follows easily that Adj is a 2-to-1 local diffeomorphism from the open set of invertible 3 × 3 symmetric matrices to the open set of 3 × 3 symmetric matrices with positive determinant. In particular, it follows that γ is a 2-to-1 local diffeomorphism from the open set of non-degenerate quadratic forms on V to the open set of quadratic forms on Λ2 (V ) with positive determinant. Thus, if h is a non-degenerate quadratic form on V , then the differential of γ at h is an isomorphism and hence is surjective. In the case r = 2, we may let w1 , w2 be an orthonormal basis of W . If we write h ∈ W ⊗ S 2 (V ∗ ) in the form h = w1 ⊗ h1 + w2 ⊗ h2 where h1 , h2 belong to S 2 (V ∗ ), then γ(h) = γ(h1 ) + γ(h2 ). Since any non-zero quadratic form q on Λ2 (V ) can be written as a sum q = q1 + q2 where each qi has positive determinant, it follows that when r = 2, the map γ : W ⊗ S 2 (V ∗ ) → K(V ) is surjective and, moreover, if O ⊂ W ⊗ S 2 (V ∗ ) is the open set on which the differential of γ is surjective, then γ(O) contains all non-zero elements of K(V ). In fact, O can be characterized as the set of h ∈ W ⊗ S 2 (V ∗ ) with the property that |h| contains a non-degenerate quadratic form. If h ∈ W ⊗ S 2 (V ∗ ) satisfies γ(h) = 0, then it is not difficult to see that h can be written in the form h = w1 ⊗ (x1 )2 + w2 ⊗ (x2 )2 where w1 , w2 form an orthonormal basis of W and x1 , x2 are elements of V ∗ . Since such an h does not belong to O, it follows that γ(O) = K(V ) − {0}.

260

VII. Examples

Let us see how this analysis is reflected in the corresponding isometric embedding problems. First, the case r = 1 corresponds to the problem of isometrically embedding N 3 into E4 . For simplicity, let us consider the case where the Riemann curvature tensor is non-degenerate at every x ∈ N when regarded as a quadratic form on Λ2 (Tx N ). This corresponds to the assumption that R : F → K3 ∼ = S 2 (Λ2 (R3 )) satisfies det(R(f)) = 0 for all f ∈ F. If det(R(f)) < 0 for all f ∈ F, then by our above discussion, the function T (f, f , h) = γ(h) − R(f) never vanishes on F × F(E4 ) × (R1 ⊗ S 2 (R3 )). Thus Z is empty and hence there do not exist any local isometric embeddings of N into E4 . This situation occurs, for example, when all of the sectional curvatures of N are negative since then, R is everywhere negative definite. On the other hand, if det(R(f)) > 0 for all f ∈ F, then the above discussion shows that for each f ∈ F, there exist exactly two solutions of the equation γ(h) = R(f), each of which is the negative of the other. Let us write h(f) for the unique element of S 2 (R3 ) which has positive determinant and which satisfies γ(h(f)) = R(f). Then Z consists of the triples (f, f , ±h(f)). Since the 6 components of γ have linearly independent differentials along Z, it follows that Z ∗ = Z = Z. When we restrict to Z ∗ , the symbol relations (13) may be solved for the πij in the form (102) πij ≡ (det(R))−3/2 Rijmωm mod I. m

Here, the functions Rijk = Rjik are some universal polynomial expressions which are of degree 4 in the components Rijkl and linear in the components Rijklm. Referring to the structure equations (12), we see that the torsion of the system (I, Ω+ ) on Z ∗ vanishes if and only if Rijk = Rikj . Let us define Tijk = Rijk − Rikj . Due to the symmetry Rijk = Rjik , there are only 8 independent functions among the Tijk . It is easy to see that these functions are the components of a well-defined tensor T of rank 8 on N . On any neighborhood of a point where T is non-zero, there is no local isometric embedding of N into E4 . On the other hand, if T vanishes identically (and det(R) > 0) then the system I is a Frobenius system on Z ∗ . Since it is clear that the group of rigid motions of E4 acts transitively on the space of leaves of I, it follows that there exist local isometric embeddings of N into E4 and that they are unique up to rigid motions. Incidentally, even though T has 8 components, it can be shown that T satisfies a set of 3 linear conditions whose coefficients are linear in the components of R. Thus, the condition T = 0 is actually only 5 conditions on a metric g which satisfies det(R) > 0. In fact, a little algebra shows that these 5 conditions can be expressed as the vanishing of a tensor of rank 5 whose components are cubic polynomials which are quadratic in the components of R and linear in the components of ∇R. We now turn to the case r = 2. Let us begin by assuming that the metric g has the property that its Riemann curvature tensor R does not vanish identically at any point of N . As we have seen, when r = 2, the map γ : W ⊗ S 2 (V ∗ ) → K(V ) restricts to become a surjective submersion γ : O → K(V ) − {0}. Let O∗ ⊂ O be the dense open subset consisting of those h ∈ O so that |h| has dimension 2. It then follows that ZO∗ = Z ∩ (F × F(E5 ) × O∗ ) consists entirely of ordinary zeros of the collection T and that ZO∗ ⊂ Z ∗ . Since R is non-vanishing on F , it is not difficult to see that ZO∗ is a dense open subset of Z ∗ and that its projection onto F × F(E5 ) is a surjective submersion. If the sectional curvature of g is everywhere negative, then it can be shown that, in fact, Z = ZO∗ = Z ∗ .


261

We now proceed to investigate the symbol and torsion of the system (I, Ω+ ) on ZO∗ . Of course, the symbol relations (95) play the crucial role. Suppose that (f, f , h) ∈ ZO∗ . Since h ∈ O∗ , it follows easily that there exists an orthonormal basis w1 , w2 of R2 = W so that, when we expand h in the form h = w1 ⊗ h1 + w2 ⊗ h2 , the quadratic form h2 is non-degenerate and not positive definite. Let v ∈ R3 be a null vector for h2 which is not a null vector for h1 . Since the results of computing the ranks of polar equations, etc. for I will be the same at all points on a given O(2) × O(3)-orbit in ZO∗ , we may assume that v is a multiple of the first element basis of R3 . It follows that when we expand h of thei standard j in the form h = i,j hij x ◦ x , the vector h11 is a non-zero multiple of w1 and the quadratic form h2 = w2 · h is non-degenerate. We now want to show that the symbol relations (95) and πij = πji at (f, f , h) ∈ ZO∗ imply that s1 = 6 and sp = 0 for all p > 1. To see this, it suffices to show that the reduced symbol relations allow us to express all of the components of π22 , π23 , and π33 as linear combinations of the forms in J (= span of the forms in I and the forms ωi ) and the components of π11 , π12, and π13. If we let K denote the span of the forms in J and the components of π11, π12 and π13 , then the symbol relations (13) imply h11 · π22 (103)

h11 · π23

h11 · π33 −h13 · π22 + h12 · π23 −h13 · π23 + h12 · π33 h33 · π22 − 2h23 · π23 + h22 · π33

≡ ≡ ≡ ≡ ≡ ≡

⎫ 0⎪ ⎪ 0⎪ ⎪ ⎪ ⎬ 0 0⎪ ⎪ ⎪ 0⎪ ⎪ ⎭ 0

mod K

(For example, the fourth relation is obtained by setting (i, j, k, l) = (1, 2, 2, 3) in (95).) The first three relations in (103) show that the w1 -components of π22, π23 , and π33 belong to K. Due to the non-degeneracy of h2 , the last three relations in (103) then show that the w2 -components of π22 , π23 , and π33 also belong to K. Since there are at least 6 linearly independent components among the πij , it follows that all of the components of π11 , π12 , and π13 are linearly independent. Thus, s1 = 6 and sp = 0 for all p > 1, as claimed. It follows by Cartan’s test that the space of integral elements of (I, Ω+ ) based at any point of ZO∗ is of dimension at most 6. We are now going to show that if the torsion of (I, Ω+ ) vanishes at (f, f , h) ∈ ZO∗ , then there exists a 6-parameter family of integral elements of (I, Ω+ ) based at (f, f , h) ∈ ZO∗ . To do this, we must show that the homogeneous system (104)

hik · hjlm − hil · hjkm + hjl · hikm − hjk · hilm = 0

for the R2 -valued unknowns hijk = hjik = hikj has a 6-parameter family of solutions. Note that by Cartan’s test, this system of equations cannot have more than 6 linearly independent solutions. We shall show that, for an open subset of h ∈ O∗ , (104) has exactly 6 linearly independent solutions. By the principle of specialization and the upper bound on the space of solutions of (104) given by Cartan’s test, this will imply that (104) has exactly 6 linearly independent solutions for all h ∈ O∗ . Our argument will use the GL(V )-equivariance of the equations (104). For an open set of h ∈ R2 ⊗ S 3 (R3 ), the corresponding two dimensional space of quadratic

262

VII. Examples

forms |h| has the property that it contains no perfect squares (i.e., rank 1 quadratic forms) and the property that there exists a basis of R3 in which the elements of |h| are diagonalized. Let x1 , x2 , x3 be linear coordinates on R3 so that such an h can be written in the form h = h11 (x1 )2 + h22 (x2 )2 + h33 (x3 )2 where the vectors hii lie in R2 . Note that the hypothesis that |h| does not contain any rank 1 quadratic forms implies that the vectors h11 , h22 , h33 are pairwise linearly independent. The equations (104) may now be written in the form

(105)

h22 · h33m h33 · h11m h11 · h22m

h11 · h23m h22 · h31m h33 · h12m − h33 · h22m − h11 · h33m − h22 · h11m

=0 =0 =0 =0 =0 =0

for all m. A priori, this is 18 equations for the 20 unknown components of the 10 vectors hijk . However, at most 14 of these equations can be linearly independent. To see this, note that if we set m = 1 (resp. 2, 3) in the first (resp. second, third) equations of (105) and use the symmetry of hijk , then we get the 3 equations hkk · h123 = 0. Due to the linear dependence of the three vectors hkk , these 3 equations are linearly dependent. Also, if we set m = 1 in the fourth equation of (105), then we see that it is a linear combination of the second equation (with m = 3) and the third equation (with m = 2). Similarly, the fifth equation with m = 2 and the sixth equation with m = 3 are linear combinations of previous equations. Thus, there are at most 14 linearly independent equations in (105). It follows that the solution space of (105) must have dimension at least 20 − 14 = 6. Since we have already seen that the solution space has dimension at most 6, we are done. Our analysis so far of the case r = 2 has shown that the homogeneous symbol relations of the system (I, Ω+ ) on ZO∗ are involutive and that the Cartan characters are s1 = 6 and sp = 0 for all p > 1. Thus, if the torsion vanishes identically, then the system (I, Ω+ ) on ZO∗ is involutive. However, in general, the torsion does not vanish identically on ZO∗ . On heuristic grounds, this is obvious since the generic metric on N 3 cannot be isometrically embedded into E5 . It is of some interest to see what the torsion of the system actually is. We claim that the unabsorbable torsion of the system (I, Ω+ ) on ZO∗ takes values in a vector space of dimension 1. To see this, note that the components of ∇R satisfy the second Bianchi identity, namely (106)

Rijklm + Rijmkl + Rijlmk = 0.

This implies that there are only 15 independent components of the covariant derivative of the Riemann curvature tensor. In other words, there is a subspace K 1 (V ) ⊂ K(V ) ⊗ V ∗ of dimension 15 in which ∇R takes values (for general n, dim K 1 (V ) = n2 (n2 − 1)(n + 2)/24, see Berger, Bryant and Griffiths [1983]). For h ∈ O∗ , define γh : W ⊗ S 3 (V ∗ ) → K 1 (V ) by the formula in indices (107)

(γh (p))ijklm = hik · pjlm − hil · pjkm + hjl · pikm − hjk · pilm


263

where p = ijk pijk xi ◦ xj ◦ xk belongs to W ⊗ S 3 (V ∗ ). We have already seen that γh has rank 14 for all h ∈ O∗ . It follows that there exists a linear functional λh : K 1 (V ) → R whose coefficients are polynomials in h of degree 14 so that ker λh = im γh . In particular, the torsion of (I, Ω+ ) is absorbable at (f, f , h) if and only if λh (∇R(f)) = 0. It can happen that λh (∇R(f)) vanishes identically on ZO∗ . For example, if the Riemannian manifold (N 3 , g) is locally symmetric, then ∇R ≡ 0, so, a fortiori, λh (∇R) ≡ 0. A more general family of metrics for which this condition holds is the family of metrics induced on the non-degenerate quadratic hypersurfaces in E4 . In fact, these examples, together with the corresponding metrics induced on the space-like portions of non-degenerate quadratic hypersurfaces in Minkowski space M4 exhaust the list of metrics with non-degenerate curvature on 3-manifolds on which this condition holds (see Berger, Bryant and Griffiths [1983]). Thus, let us assume that g has the property det(R) = 0 and that λh (∇R) ≡ 0. By our above description, such metrics are known to be real analytic in appropriate local coordinates. As we have shown, the system (I, Ω) on YO∗ = ZO∗ /O(2) × O(3) is involutive. It follows that we may apply the Cartan–K¨ ahler theorem to this system to conclude that such metrics can be locally isometrically embedded into E5 and that the general such embedding depends locally on 6 functions of 1 variable. Note that when det(R) > 0, such metrics can be isometrically embedded into E4 , but that they are rigid there. More information about these embeddings can be obtained by studying the characteristic variety of the system (I, Ω). For each integral element E of (I, Ω) the complex characteristic variety consists of 6 points. In fact, by our above computa2 ∗ tion of the complex characteristic variety ΞC H for H ∈ W ⊗ S (V ), these points are described as follows. If E is an integral element based at (x, y, l, H) ∈ YO∗ , then |H| ⊂ S 2 (E ∗ ) has dimension 2 and contains at least one non-degenerate quadratic form. In the terminology of algebraic geometry, |H| is called a pencil of quadrics. If h1 ∈ |H| is non-degenerate and h2 ∈ |H| is linearly independent from h1 , then the degenerate quadrics in |H| are of the form h2 + th1 where t is a root of the cubic equation det(h2 + th1 ) = 0. For each such root (counted with multiplicity) the quadric h2 + th1 factors in the form λ ◦ µ where λ, µ ∈ (E ∗ )C are well defined up to scalar multiples. There are six such elements of (E ∗ )C (counted with multiplicity). These give rise to the 6 points in P(E ∗ )C which constitute ΞC E. For the generic pencil of quadrics |H| ⊂ S 2 (E ∗ ), the base locus B consists of 4 points in PE C in general position except for the condition of being invariant under conjugation. Let us say that H is general if the base locus of |H| consists of 4 points. Conversely, given a conjugation-invariant set B of 4 points in PE C with no 3 on a line, there is a unique real pencil of quadrics |HB | whose elements pass through all 4 points of B. The corresponding characteristic variety ΞB ⊂ P(EC )∗ consists of the 6 lines which pass through 2 of the 4 points of B. Note that if all of the points of B are real, then all of the 6 points of ΞB are also real. If 2 of the points of B are real and the other 2 are non-real complex conjugates, then 2 of the 6 points of ΞB are real. If B consists of 2 pairs of non-real complex conjugate points, then 2 of the points of ΞB are real. In each case, note that ΞB contains at least 2 real points. Thus, it follows that the system (I, Ω) is never elliptic on YO∗ = ZO∗ /O(2) × O(3). Let us say that an immersion u : N 3 → E5 is general if the induced second fundamental form Hx is general for all x ∈ N . We have seen that general isometric

264

VII. Examples

immersions exist locally and depend on 6 functions of 1 variable for the special case of (N 3 , g) = (S 3 , can) since the canonical metric on S 3 satisfies ∇R ≡ 0. We shall now show that the integrability of the characteristic variety obstructs the existence of a global general isometric immersion of (S 3 , can) into E5 . Proposition 4.7. There does not exist a general isometric immersion u : S 3 → E5 when S 3 is given any one of the metrics induced by immersion as an ellipsoidal hyperquadric in E4 . Proof. Suppose that such a u exists. Consider the associated integral uˆ : S 3 → YO∗ of (I, Ω). Via uˆ, the characteristic variety of (I, Ω) restricts to the projectivized complexified cotangent bundle of S 3 to become a submanifold Ξuˆ ⊂ P(T ∗ S 3 )C whose base point projection Ξuˆ → S 3 makes Ξuˆ into a covering space of degree 6. By simple connectivity of S 3 , it follows that Ξuˆ is the disjoint union of 6 copies of S 3 . Moreover, the base locus Buˆ ⊂ P(T S 3 )C dual to Ξuˆ is a covering space Buˆ → S 3 of degree 4 and hence consists of the disjoint union of 4 copies of S 3 . The fiber (Buˆ )x ⊂ P(Tx S 3 )C at each point x ∈ S 3 consists of 4 (distinct) points in general position subject only to the condition of being invariant under conjugation. It follows easily that the number of real points in (Buˆ )x is the same for all x ∈ S 3 . If all of the points in (Buˆ )x are real (for all x), then due to the fact that GL(3, R) acts transitively on the set of quadruples of points in RP2 in general position, it follows easily that there exists a coframing η 1 , η 2 , η 3 on S 3 (not necessarily orthogonal) with the property that the 6 projectivized 1-forms {[η 1 ], [η 2], [η 3 + η 1 + η 2 ], [η 3 + η 1 − η 2 ], [η 3 − η 1 + η 2 ], [η 3 − η 1 − η 2 ]} are sections of Ξuˆ . By the involutivity of (I, Ω) on YO∗ , any 1-form η for which [η] is a section of Ξuˆ is integrable, i.e., η ∧ dη = 0. Applying this integrability to each of the 6 1-forms above gives 6 equations which are equivalent to the 6 relations (108)

η i ∧ dη j + η j ∧ dη i = 0

for all i and j. It is easy to show that this is equivalent to the condition that there exist a 1-form λ so that dη i = λ ∧ η i for all i. Differentiating this last relation gives the condition dλ ∧ η i = 0 for all i. Since dλ is a 2-form, it follows that dλ = 0. Due to the simple connectivity of S 3 , it follows that there exists a function l on S 3 so that λ = dl. It then follows that the coframing η˜i = e−l η i satisfies d˜ η i = 0 for all 3 i. Since S clearly does not have any closed coframing, we have a contradiction. In the case that 2 of the points in (Buˆ )x are real (for all x), then it is not difficult to see that there exists a coframing η 1 , η 2 , η 3 on S 3 (not necessarily orthogonal) with the property that the 6 projectivized 1-forms {[η 1 ], [η 2], [η 3 + iη 1 + η 2 ], [η 3 + iη 1 − η 2 ], [η 3 − iη 1 + η 2 ], [η 3 − iη 1 − η 2 ]} are sections of Ξuˆ . In the case that none of the points in (Buˆ )x are real (for all x), then it is not difficult to see that there exists a coframing η 1 , η 2 , η 3 on S 3 (not necessarily orthogonal) with the property that the 6 projectivized 1-forms {[η 1 ], [η 2], [η 3 + iη 1 + iη 2 ], [η 3 + iη 1 − iη 2 ], [η 3 − iη 1 + iη 2 ], [η 3 − iη 1 − iη 2 ]} are sections of Ξuˆ . In either case, applying the integrability of the characteristic variety shows that the equations (108) hold for this coframing. We have already seen


265

that the equations (108) lead to a contradiction when the domain of the coframing is S 3 Note that there do exist immersions of S 3 into E5 which are general. For example, if we regard E5 as the space of traceless symmetric 3 × 3 matrices and let Σ ⊂ E5 denote the set of such matrices whose eigenvalues are {1, −1, 0}, then the universal covering space of Σ is easily seen to be S 3 and it is not difficult to see that the second fundamental form of Σ is general at all points (use the fact that Σ is an orbit of SO(3) under its natural irreducible action on E5 ). Of course, the induced metric on Σ, even though homogeneous, cannot be locally isometric to any of the “quadric” metrics for which the system (I, Ω) is involutive on YO∗ . As our final example in the case r = 2, let us consider the case where the metric g is flat. Then, the Riemann curvature satisfies R ≡ 0 on F . Let us describe the set of ordinary zeros of the function T on F × F(E5 ) × (R2 ⊗ (S 2 (R3 )). Since R ≡ 0, it follows that T (f, f , h) = γ(h), and since the differential of γ is not surjective at any h ∈ R2 ⊗ S(R3 ) for which γ(h) = 0, we cannot directly apply the implicit function theorem to conclude that T −1 (0) = F × F(E5 ) × γ −1 (0) has any smooth points. Nevertheless, by our previous discussion, we can parametrize γ −1 (0) as follows. Let O(2) denote the set of orthonormal bases (w1 , w2 ) of W = R2 . Then there exists a map µ : O(2) × V ∗ × V ∗ → W ⊗ S 2 (V ∗ ) defined by µ(w1 , w2 ; x1 , x2 ) = w1 ⊗ (x1 )2 + w2 ⊗ (x2 )2 whose image is precisely γ −1 (0) ⊂ W ⊗ S 2 (V ∗ ). If we let D ⊂ R5 × R3 = V ∗ × V ∗ denote the open set of pairs (x1 , x2) of elements of V ∗ which satisfy x1 ∧ x2 = 0, then it is easy to see that the differential of µ has its maximum rank 7 precisely on the open subset O(2) × D. In fact, µ(O(2) × D) ⊂ γ −1 (0) consists of the set of h ∈ γ −1 (0) for which |h| is of dimension 2 and is a smooth submanifold of W ⊗ S 2 (V ∗ ) of dimension 7. Any point h ∈ γ −1 (0) for which |h| has dimension 1 or 0, is in the closure of µ(O(2) × D) but is not a smooth point of γ −1 (0). Thus, in order to see that F × F(E5 ) × µ(O(2) × D) is the space of ordinary zeros of the collection T , it suffices to show that the differentials of the functions in T contain at least 5 linearly independent 1-forms at every point of F × F(E5 ) × µ(O(2) × D). To see this, note that we may take advantage of the GL(V )-equivariance of the map γ. Let h ∈ µ(O(2) × D) and write h in the form h = w1 ⊗ (x1 )2 + w2 ⊗ (x2 )2 . (The elements x1 , x2 are not necessarily an orthonormal pair in V .) Making the appropriate GL(V ) change of basis for the 1-forms on F × F(E5 ) × (W ⊗ S 2 (V ∗ )), we can assume that h11 and h22 are an orthonormal basis of W and that all other hij are zero. Then a basis for the 1-forms τijkl at (f, f , h) can be expressed in the form (109)

τ2323 = h22 · π33,

τ3131 = h11 · π33

τ1212 = h11 · π22 + h22 · π11 , τ1223 = −h22 · π13 ,

τ3112 = −h11 · π23 τ2331 = 0.

It follows that there exist 5 linearly independent 1-forms among the τijkl at (f, f , h), as we wished to show. Thus, Z = Z ∗ = F × F(E5 ) × µ(O(2) × D). We are now going to show that the differential system (I, Ω+) is involutive on Z ∗ with Cartan characters s1 = 6, s2 = 1, and sp = 0 for all p > 2. First, note that since R ≡ 0, we must have ∇R ≡ 0 as well. Thus, the torsion of the differential

266

VII. Examples

system vanishes identically on Z ∗ . Examining the symbol relations, i.e., the right hand sides of (109), we see by inspection that there is a flag contained in each integral element which has the characters s1 = 6, s2 = 1, and sp = 0 for all p > 2. Direct calculation the the space of integral elements at each point of Z ∗ shows that there exists an 8-parameter family of integral elements at every point of Z ∗ . Thus, by Cartan’s test, it follows that the system is in involution, as claimed. Note that applying Proposition 4.1 shows that the characteristic variety ΞH of an element H = w1 ⊗ (x1 )2 + w2 ⊗ (x2 )2 consists of the line [ξ1 x1 + ξ2 x2 ] in PV ∗ . Note that ΞH has degree and dimension 1 in accordance with the general theory. It follows that the submanifolds N 3 ⊂ E5 on which the induced metric is flat and whose second fundamental forms have rank 2 at each point depend on one function of two variables (in Cartan’s terminology). For further examples of isometric embedding for special metrics in codimensions below the natural embedding codimension, the reader may consult the aforementioned paper (Berger, Bryant and Griffiths [1983]) and its references to the work of Cartan. In particular, Cartan’s study of γ −1 (0) ⊂ Kn constitutes his theory of “exteriorly orthogonal quadratic forms” which he used to study the isometric embeddings of En into E2n and Hn into E2n−1 (here, Hn denotes hyperbolic n-space). In the latter problem, the analog of the equations (108), which are consequences of the integrability of the characteristic variety, can be used to prove the existence of “generalized Tschebysheff coordinates” which can be associated to any isometric embedding of Hn into E2n−1 (see Moore [1972]).

Introduction

267

CHAPTER VIII APPLICATIONS OF COMMUTATIVE ALGEBRA AND ALGEBRAIC GEOMETRY TO THE STUDY OF EXTERIOR DIFFERENTIAL SYSTEMS

A linear Pfaffian differential system on a manifold M is given by sub-bundles I ⊂ J ⊂ T ∗ (M ) such that dI ⊂ {J} where {J} ⊂ Ω∗ (M ) is the algebraic ideal generated by the sections of J. Setting L = J/I it follows that the exterior derivative induces a bundle mapping (cf. Section 5 of Chapter IV) δ : I → (T ∗ (M )/J) ⊗ L. Dualizing and using (T ∗ (M )/J)∗ ∼ = J ⊥ , this is equivalent to a bundle mapping (1)

π : J ⊥ → I ∗ ⊗ L.

Locally, this mapping is given by the tableau matrix π as discussed in Chapter IV. Much of the discussion in the preceeding chapters has centered around fibrewise constructions, such as the symbol and characteristic variety, associated to the mapping (1). In this chapter we will isolate and considerably extend these discussions. Given vector spaces W and V , a tableau has been defined to be a linear subspace A ⊂ W ⊗ V ∗, and the associated symbol has been defined to be B = A⊥ ⊂ W ∗ ⊗ V. As explained in Chapter IV, the first prolongation A(1) ⊂ W ⊗ S 2 V ∗ is defined by A(1) = {P ∈ W ⊗ S 2 V ∗ : v

P ∈ A for all v ∈ V },

and A is said to be involutive in case equality holds in the inequality given by Cartan’s test (cf. Proposition 3.6 in Chapter IV). This chapter will be an algebraic study of involutive tableau. Three of the basic results are stated in (2.4), (2.5) and (3.3) below. One of the most useful exterior algebra facts is the Cartan lemma, and in Section 2 we begin with a generalization (Proposition 2.1) of this lemma that plays a

268

VIII. Applications of Commutative Algebra

critical role in the theory of exterior differential systems. This lemma leads naturally to the definition of the Spencer cohomology groups H k,q (A) of a tableau. The crucial role played by these groups in the development of the subject of overdetermined P.D.E. systems will be explained in Chapters IX and X. For us the main fact will be the characterization of involutive tableau as those for which all H k,q (A) = 0, k ≥ 1 (cf. the discussion in Section 1 of Chapter X). This result, which among other things implies that the prolongation of an involutive tableau is involutive, will be proved in Sections 2 and 3 below. Actually, we have chosen to also give a direct proof of the result that “A involutive ⇒ A(1) involutive” in Section 2, as among other things it shows how one is led naturally to the Spencer cohomology groups by purely differential system considerations. For another example of how cohomology naturally arises, it has been remarked in Chapter IV and will proved below that the torsion of a linear Pfaffian differential system lies in the family of vector spaces H 0,2 (Ax ), where Ax is the tableau lying over x ∈ M . In Section 3 we dualize Cartan’s test for involution. This is done by introducing a graded module MA naturally associated to a tableau A, and the condition that A be involutive is seen to be that MA admit a quasi-regular sequence. The latter condition is, by more or less standard commutative algebra, equivalent to the vanishing result Hk,q (MA ) = 0, k ≥ 1 for the Koszul homology groups of the module MA . These Koszul homology groups then turn out to be dual to the Spencer cohomology groups, thus establishing the equivalence of involutivity and the vanishing of Spencer cohomology. At the end of Section 3, this characterization of involutivity is used to prove the fact that, given a tableau A, there is a q0 such that the prolongations A(q) of A are involutive for q ≥ q0 , a result used in the proof of the Cartan–Kuranishi theorem. In fact, we prove the stronger result that q0 depends only on the sequence of numbers dim A(q) ; this requires a localization argument and is related to the construction of Hilbert schemes in algebraic geometry. In Section 4, we further pursue the use of Koszul homology by showing that the above vanishing result leads to a natural definition of what is meant by an involutive module, and then it is shown that involutive modules have canonical free resolutions where the maps are homogeneous of degree one. This translates into statements such as: the compatibility equations for an involutive, overdetermined linear P.D.E. system are of first order (see Theorem 1.8, Chapter X), and so forth. The results in this section will be used in Section 6 when Guillemin’s normal form is discussed. In Section 5 we introduce what amounts to the micro-localization of a linear Pfaffian differential system. The concepts an involutive sheaf and of the characteristic sheaf of a tableau are introduced, and more or less standard algebro-geometric results are used to prove a number of results about these, culminating in the proof of Theorem 3.15 in Chapter V and the proof of Proposition 3.10 below, which is the essential algebraic step in the proof of Theorem 3.1 in Chapter VI. The formal introduction of homological methods is due to Spencer [1961]. Most all of the results in this chapter were found in the early and middle 1960’s and are due to Guillemin, Quillen, Spencer and Sternberg with crucial input coming from Mumford and Serre. In addition to Spencer’s paper cited above, we would like to call attention to Singer and Sternberg [1965], Guillemin and Sternberg [1964],

§1. Involutive Tableaux

269

Quillen [1964] and Guillemin [1968] where much of the theory first appeared. §1. Involutive Tableaux. We begin by introducing some notations. Let W be a vector space with basis {wa }, V a vector space with basis {vi } and dual basis {xi }, S q V ∗ the q th symmetric product of V ∗ , and W ⊗ S q V ∗ the W -valued polynomials of the form P = PIa wa ⊗ xI where I = (i1 , . . . , iq ) runs over multi-indices of length q and xI = xi1 . . . xiq . By ∂P = vi ∂xi

P

we mean the formal derivative treating W as constants. We repeat and generalize some definitions from Chapter IV. Definition 1.1. i) A linear subspace A ⊂ W ⊗ V ∗ will be called a tableau. More generally, a subspace A ⊂ W ⊗ S p+1 V ∗ will be called a tableau of order p. ii) Given a tableau of order p, we inductively define the q th prolongation A(q) ⊂ W ⊗ S p+q+1 V ∗ by A(0) = A and A(q) = {P ∈ W ⊗ S p+q+1 V ∗ :

∂P ∈ A(q−1) for all i}. ∂xi

To motivate this definition in the case p = 0 of an ordinary tableau, we suppose that the linear Pfaffian system (I, Ω) has no integrability conditions, and we denote its first prolongation by (I (1) , Ω) on the manifold M (1). If Ax denotes the tableau of (I, Ω) at a typical point x ∈ M , then the fibre of π : M (1) → M (1)

over x is an affine linear space whose associated vector space is Ax (cf. (125) in Chapter IV). Now (I (1), Ω) is again a linear Pfaffian differential system whose tableau at a (1) point y with π(y) = x is Ax . Assuming again that there are no integrability conditions, the second prolongation is a linear Pfaffian differential system (I (2) , Ω) over the manifold M (2) where π (1) : M (2) → M (1) (2)

is a family of affine linear spaces whose associated linear space z ∈ M (2) is Ax where (π (1) ◦ π)(z) = x.

270


In summary, in the absence of integrability conditions the fibres of the prolongations π (q−1) : M (q) → M (q−1) are affine linear spaces whose associated vector spaces are the prolongations A(q) of a tableau A. More precisely, if x0 , x1 , x2 , . . . is a sequence of points xq ∈ M (q) with (q) π (q−1) (xq ) = xq−1 , then Axq = Ax0 . q ∗ the Returning to the general discussion, if we denote by S + V ∗ = q≥1 S V ∗ q ∗ maximal ideal in the polynomial algebra SV = q≥1 S V and define the total prolongation to be A = q≥0 A(q) ⊂ W ⊗ S + V ∗ , then it is clear that A is the largest graded subspace of W ⊗ S + V ∗ that is closed under differentiation and satisfies

(2)

A ∩ (W ⊗ S p+1 V ∗ ) = A, A ∩ (W ⊗ S q V ∗ ) = 0 for q ≤ p. We note that the grading on A is shifted by p + 1 from that on S + V ∗ ; that is A(q) ⊂ W ⊗ S p+q+1 V ∗ . Another useful characterization of prolongations is based on the following observation: If we consider each of S p+1 V ∗ ⊗ S q V ∗ , S p V ∗ ⊗ S q+1 V ∗ , and S p+q+1 V ∗ as subspaces of ⊗p+q+1 V ∗ , then (3)

(S p+1 V ∗ ⊗ S q V ∗ ) ∩ (S p V ∗ ⊗ S q+1 V ∗ ) = S p+q+1 V ∗ ,

and consequently A(q) = (A ⊗ S q V ∗ ) ∩ (W ⊗ S p+q+1 V ∗ ).

(4)

We shall now recall and extend the concept of involution for a tableau. Referring to a definition from Chapter IV we have the subspaces (q)

Ai

= {P ∈ A(q) :

∂P ∂P = ···= = 0}, ∂x1 ∂xi

where x1 , . . . , xn is assumed to be a general basis of V ∗ . Remark that (q)

(5)

Ai

= (Ai )(q)

(0)

where Ai = Ai , and that (q)

(q) . A(q) n = (0), A0 = A

The proof of Proposition 3.6 in Chapter IV works equally for a general tableau and gives:

§1. Involutive Tableaux

271

Proposition 1.2. We have dim A(1) ≤ dim A + dim A1 + · · · + dim An−1 with equality holding if, and only if, the maps ∂ (1) : Ai−1 → Ai−1 ∂xi

(6) are surjective for i = 1, . . . , n.

If we define the characters s1 , . . . , sn of a tableau A by (cf. Definition 3.5 in Chapter IV) s1 + · · · + sk = dim A − dim Ak , then the inequality in Proposition 1.2 is dim A(1) ≤ s1 + 2s2 + · · · + nsn . Here and throughout this chapter we have dropped the primes since these are the only characters which we shall consider, and we shall also set s = dim W instead of using s0 as was done in Chapter IV. Definition 1.3. A tableau of order p is involutive if dim A(1) = dim A + dim A1 + · · · + dim An−1 . Being involutive is equivalent to the maps (6) being surjective. This is equivalent to the equality dim A(1) = s1 + 2s2 + · · · + nsn in Cartan’s test, which agrees with Definition 3.7 in Chapter IV. The following are the basic properties of involutive tableaux: (7)

Every prolongation of an involutive tableau is involutive.

(8)

If A is any tableau, then there is a q0 such that the prolongations A(q) are involutive for q ≥ q0 .

These will be proven later in this chapter. For the moment we shall use (7) to prove the relations (which we have already encountered in Chapter III) (1)

sn = sn (1) sn−1 = sn + sn−1 .. .. . . (1) s1 = sn + · · · + s1

(9)

(1)

where the sk are the characters of A(1) . As a corollary we have that (10)

The character l, i.e., the largest l such that sl = 0, and the Cartan integer sl are invariant under prolongation of an involutive tableau.

272


Proof of (9). By definition (1)

(1)

(1)

s1 + · · · + sk = dim A(1) − dim Ak . On the other hand, by the surjectivity of the maps (6) all the tableaux Ak are involutive. Hence (1)

dim Ak = dim Ak + · · · + dim An−1 (1)

⇒ dim A(1) − dim Ak = dim A + · · · + dim Ak−1 (1)

(1)

⇒s1 + · · · + sk = dim A + · · · + dim Ak−1 (1)

⇒sk = dim Ak−1 = sk + · · · + sn . Example 1.4. Let A ⊂ W ⊗ V ∗ be a tableau, B = A⊥ ⊂ W ∗ ⊗ V be the symbol relations, and B λ = Baλi wa∗ ⊗ vi a basis for B. In the jet space J 1 (V, W ) with coordinates (xi , z a , pai) let M be defined by the equations Baλi pai = 0. The restriction to M of the contact system

dz 1 − pai dxi = 0 dx1 ∧ · · · ∧ dxn = 0

corresponds to the linear, homogeneous constant coefficient P.D.E. system Baλi

(11)

∂z a (x) = 0. ∂xi

We observe that The total prolongation A is the space of formal power series solutions, with zero constant term, to (11). This is immediate from a q+1 a ∂q z (x) λi ∂z (x) λi ∂ (B ) = B . a a I i I ∂x ∂x ∂x ∂xi

The involutivity of the tableau A is equivalent to the involutivity of the Pfaffian differential system associated to (11)—cf. Proposition 3.8 in Chapter IV. More generally, let A ⊂ W ⊗ S p+1 V ∗ be a tableau of order p. We may identify ∗ W ⊗ SV with the constant coefficient differential operators on W ⊗ SV ; thus (wa∗ ⊗ vI )(PJb wb ⊗ xJ ) =

∂ (P a xJ ). ∂xI J

§2. The Cartan–Poincaré Lemma, Spencer Cohomology

273

Giving A is equivalent to giving its annihilator B = A⊥ ⊂ W ∗ ⊗ S p+1 V , so that a tableau of order p gives a linear, homogeneous constant coefficient P.D.E. system of order p + 1 and vice versa. Roughly speaking, a general linear Pfaffian differential system has the two aspects consisting of its tableau and torsion. To be involutive means that i) for each x, the tableau is involutive (i.e., the constant coefficient system corresponding to x—like freezing the leading coefficients in a P.D.E.—is involutive); and ii) the torsion vanishes (i.e., the integralibility conditions are satisfied—cf. Theorem 5.16 in Chapter IV). As we shall explain below, the tableau of (I, Ω) influences both the tableau and torsion of (I (1), Ω). §2. The Cartan–Poincar´ e Lemma, Spencer Cohomology. One of the most useful facts in exterior algebra is the familiar Cartan Lemma: Let V be a vector space and suppose there is a quadratic relation wi ∧ vi = 0, wi , vi ∈ V, i

where the vi are linearly independent. Then wi = j aij vj where . aij = aji We shall give a generalization of the result that plays a crucial role in our theory of exterior differential systems. Let U and V be vector spaces and suppose given a linear map Ω:U →V with adjoint We set

Ω∗ : V ∗ → U ∗ . C p,q = S p V ∗ ⊗ Λq U ∗

and define a boundary operator δΩ : C p,q → C p−1,q+1 by the rule (12)

δΩ (v1∗ ⊗ · · · ⊗ vp∗ ⊗ ψ) =

v1∗ ⊗ · · · ⊗ vˆα∗ ⊗ · · · ⊗ vp∗ ⊗ Ω∗ (vα∗ ) ∧ ψ

α 2 where v1∗ , . . . , vp∗ ∈ V ∗ and ψ ∈ Λq U ∗ . It is clear that δΩ = 0, and we denote the resulting cohomology by δ

Ω H p,q (Ω) = ker{C p,q −→ C p−1,q+1 }/δΩ C p+1,q−1 .

274


Proposition 2.1 (The Cartan–Poincaré lemma). We have the isomorphism H p,q (Ω) ∼ = S p (ker Ω∗ ) ⊗ Λq (coker Ω∗ ). We will give the proof in two steps. Step One. Suppose that Ω is an isomorphism and use it to identify U with V . If we choose linear coordinates x1 , . . . , xn on U , the elements in C p,q are ϕ= ϕi1 ...iq (x)dxi1 ∧ · · · ∧ dxiq i1 0) that satisfy the conditions ϕi1 ...iq is symmetric in i1 , . . . , iq (19) i i ϕi1 ...iq−1 i ∧ ω = 0. Then there exist ψj1 ...jq+1 ∈ Λr−1 U ∗ that satisfy

ψj1 ...jq+1 is symmetric in j1 , . . . , jq+1 j i ψj1 ...jq j ∧ ω = ϕj1 ...jq .

(20)

Proof. In this case V = Rn and Ω=U →V given by Ω(u) = (ω1 (u), . . . , ωn (u)),

u ∈ U,

is surjective. In particular H q,r (Ω) = 0 when r > 0.

(21) The conditions (19) are

ϕ ∈ S q V ∗ ⊗ Λr U ∗ δΩ (ϕ) = 0,

and the conditions (20) are

ψ ∈ S q+1 V ∗ ⊗ Λr−1 U ∗ δΩ (ψ) = ϕ.

Thus the corollary is equivalent to (21).

When r = q = 1, this corollary is the usual Cartan lemma. When Ω is an isomorphism, the Cartan–Poincaré lemma is the Poincaré lemma for polynomial differential forms. We shall give a variant of this discussion of the Cartan–Poincaré lemma. Let V, W be vector spaces and set C k,q = W ⊗ S k V ∗ ⊗ Λq V ∗ . Choosing bases {wa} for W and {xi } for V ∗ we may think of ϕ ∈ C k,q as a W -valued polynomial differential form ϕ=

wa ⊗ ϕaIJ xI dxJ .

|I|=k, |J|=q

We define δ : C k,q → C k−1,q+1


277

to be the usual exterior differentiation treating the wa as constants. Denoting the resulting cohomology by H k,q we have from (13) (22)

H

k,q

=

W

k=q=0

0

otherwise.

Now let A ⊂ W ⊗ V ∗ be a tableau with prolongations A(p) ⊂ W ⊗ S p+1 V ∗ , and define C k,q (A) ⊂ C k,q by C k,q (A) =

(23)

A(k−1) ⊗ Λq V ∗ W ⊗ Λq V ∗

k≥1 k = 0.

The defining property of prolongations given by equation (2) in §1 above implies that δ : C k,q (A) → C k−1,q+1(A), and we denote by H k,q (A) the resulting cohomology: (24)

H k,q (A) = ker{δ : C k,q (A) → C k−1,q+1 (A)}/δC k+1,q−1(A).

Definition 2.3. The H k,q (A) are the Spencer cohomology groups associated to the tableau A. With the correspondence in notation A(p) ↔ gk+p, this definition coincides with that in Chapter IX. For us their importance resides in the following result, the first half of which will be proved in a moment and the remainder in §3. Theorem 2.4. If A is involutive, then H k,q (A) = 0

k ≥ 1, q ≥ 0.

The converse is also true. In general, if A ⊂ W ⊗ S p+1 V ∗ is a tableau of order p we define C k,q (A) ⊂ C k+p,q by (25)

C k,q (A) =

A(k−1) ⊗ Λq V ∗ W ⊗ S p V ∗ ⊗ Λq V ∗

k≥1 k = 0.

This agrees with (23) when p = 0, which is the case of an ordinary tableau. Again by the defining property of prolongations we may define the Spencer cohomology groups by the same formula (24). We will now prove the following proposition, the first statement of which gives one half of Theorem 2.4, and the second statement of which gives a result pertaining to Chapter VI. The complete proof of Theorem 2.4 will be given later.

278


Proposition 2.5. Let A be a tableau of order p. (i) If A is involutive, then H k,q (A) = 0 for k ≥ 1, q ≥ 0. (ii) If A is involutive, then the prolongations A(q) are involutive for q ≥ 1. We note that (ii) follows from (26) below and Theorem 2.4; however, the proof of Theorem 2.4 is somewhat lengthy and so we will first give a direct proof of (ii). As an application of Proposition 2.5, we let A ⊂ W ⊗ V ∗ be an ordinary tableau and picture its cohomology as coming from the diagram A(2) A(1)

A W

A(2) ⊗ V ∗ A(1) ⊗ V ∗ A⊗V∗ W ⊗V∗

A(2) ⊗ Λ2 V ∗ A(1) ⊗ Λ2 V ∗ A ⊗ Λ2 V ∗ W ⊗ Λ2 V ∗ .

For the 1st prolongation its Spencer cohomology comes from the diagram A(3)

A(2)

A(1) W ⊗V∗

A(3) ⊗ V ∗ A(2) ⊗ V ∗ A(1) ⊗ V ∗ W ⊗V∗ ⊗V∗

A(3) ⊗ Λ2 V ∗ A(2) ⊗ Λ2 V ∗ A(1) ⊗ Λ2 V ∗ W ⊗ V ∗ ⊗ Λ2 V ∗ .

Comparing these two it is clear that (26)

H k,q (A(1)) ∼ = H k+1,q (A),

k ≥ 1.

In particular, we have from Proposition 2.5 above: (27)

If A is involutive, then H k,q (A(1) ) = (0),

k ≥ 1.

Returning to our general discussion, the proof of Proposition 2.5 will follow from the two assertions (28)

A involutive ⇒ H p,q (A) = 0 for p ≥ 1, q ≥ 0.

(29)

H p,q (A) = 0 for p ≥ 1 and statement (ii) in Proposition 2.5 when dim V ≤ n − 1 together imply (ii) in Proposition 2.5 when dim V = n.

Our proof of (28) is due to Sternberg, and the idea is this: Using the surjectivity of the maps (6) in §1 above, a standard proof of the Poincaré lemma carries over


279

verbatim to give H p,q (A) = 0 for p ≥ 1, q ≥ 1 (the case p ≥ 1, q = 0 must be treated separately and will be left to the reader). In fact, the exposition will be clearer if we just give this standard inductive proof of the usual Poincaré lemma and leave it for the reader to simply observe that the argument also establishes (29). Let U be the closed cube {x ∈ Rn : |xi | ≤ 1} and let ϕ ∈ Ωq (U ) be a closed C ∞ q-form (the coefficients of ϕ are assumed smooth in a neighborhood of U ). If q ≥ 1 we want to find η ∈ Ωq−1 (U ) satisfying dη = ϕ. Suppose that ϕ involves only the differentials dx1 , . . . , dxk . The construction of η will be by descending induction on k. Namely, we will inductively find ηk such that ϕ − dηk involves only dx1 , . . . , dxk−1. Then η = η1 will be our required form. To find ηk we write ϕ = ϕ + ϕ ∧ dxk where ϕ , ϕ involve only dx1 , . . . , dxk−1. From 0 = dϕ = dϕ + dϕ ∧ dxk , by looking at the coefficients of dxk ∧ dxl where l > k we see that ∂ϕ = 0, ∂xl

l > k.

Here, the derivatives of a form mean the derivatives of its coefficients. By elementary calculus, we may find a (q − 1)-form ηk involving only dx1 , . . . , dxk−1 and satisfying ∂η k = ϕ ∂xk (30) ∂ηk for l > k. ∂xl = 0 In other words, if Ck∞ (U ) are the C ∞ functions in U that only depend on x1 , . . . , xk , then the mappings ∂/∂xk : Ck∞ (U ) → Ck∞(U )

(31)

are surjective. In fact, given f ∈ Ck∞(U ) the function

xk

g(x1 , . . . , xk ) =

f(x1 , . . . , xk−1, t)dt −1

Ck∞ (U )

satisfies g ∈ Now consider

k

and ∂g/∂x = f. We note the similarity between (31) and (6).

ψ = ϕ − dηk = ϕ − ϕ ∧ dxk + terms involving only dx1 , . . . , dxk−1. By (30) the form ψ involves only dx1 , . . . , dxk−1, and finding it completes the induction step in our proof of the usual Poincaré lemma.

280


As suggested above, the proof of (28) is the same with (6) in §1 replacing (31). We now turn to the proof of (ii) in Proposition 2.5. The result is due to Cartan; a direct proof is given in Singer and Sternberg [1965] and, as we have noted, the result follows in a general way from (26) and Theorem 2.4. We shall give the argument for an ordinary tableau A ⊂ W ⊗ V ∗ , the proof for a general tableau being essentially the same. Thus we must show that: (32)

A involutive ⇒ A(1) involutive.

By our inductive strategy (29) we may assume (i) that (32) is true when dim V ≤ n − 1; (ii) that (i) in Proposition 2.5 is true when dim V ≤ n. We must then prove (32) when dim V = n. By definition and Proposition 1.2 above we must show that there exists a basis x1 , . . . , xn for V ∗ such that the maps (33i )

(2)

(1)

∂/∂xi : Ai−1 → Ai−1

are surjective for i = 1, . . . , n. By the involutivity of A there exists a basis such that the maps (34)

(1)

∂/∂xi : Ai−1 → Ai−1 ,

i = 1, . . . , n,

are all surjective, and this is the basis for V ∗ we shall use. Next we note that A1 is itself an involutive tableau in n − 1 variables. By our induction assumption we may then assume that the maps (33i ) are surjective for i = 2, . . . , n. It remains to prove that ∂/∂x1 : A(2) → A(1) is surjective. Let Q ∈ A(1) ⊂ W ⊗ S 2 V ∗ . We want to find P ∈ A(2) ⊂ W ⊗ S 3 V ∗ satisfying (35)

∂P = Q. ∂x1

The idea is to use the surjectivity of (34) to solve for the derivatives of this equation. Then the vanishing of cohomology will allow us to “integrate” this solution. Thus consider ∂Q ∈ A ⊂ W ⊗ V ∗. ∂xj By (34) when i = 1 we may find Tj ∈ A(1) with (36) Consider the 1-form

∂Q ∂Tj = . ∂x1 ∂xj T = Tj dxj ∈ A(1) ⊗ Λ1 V ∗ .


281

Its exterior derivative satisfies dT ∈ A ⊗ Λ2 V ∗ ∂ ∂T = d2 Q = 0, (dT ) = d ∂x1 ∂x1 where the derivatives of a form mean the derivatives of its coefficients. Denote by ϕ the restriction of a form whose coefficients are functions of x2 , . . . , xn to the subspace x1 = 0 (thus ϕ → ϕ means to set dx1 = 0). Then, since dϕ = dϕ, d(dT ) = 0. We may thus consider dT as a class in H 1,2 (A1 ) = 0 by our induction assumption (ii). Consequently, we may find S=

n

Sα (x2 , . . . , xn )dxα = S

α=2 (1)

with Sα ∈ A1

and dT = dS.

It follows that d(T − S) = dx1 ∧ U

(37) where U=

n

Uα (x2 , . . . , xn)dxα = U ,

Uα ∈ A1 .

α=2

Taking exterior derivatives of both sides of (37) gives 0 = dx1 ∧ dU. It follows that U ∈ H 1,1 (A1 ) = 0, again by our induction assumption. Thus U = −dW,

(1)

where W ∈ A1

⇒d(T − S) = d(W dx1 ). Setting R = S + W dx1 =

n

Rj (x2 , . . . , xn)dxj ,

i=1

where Rj ∈

(1) A1 ,

we see that d(T − R) = 0.

282


Then, by the usual Poincaré lemma for homogeneous polynomials differential forms there exists P such that (i) T − R = dP (ii) P ∈ W ⊗ S 3 V ∗ . The first equation implies that P ∈ A(2) , and we claim that (35) is satisfied. In fact ∂ ∂xj

∂P −Q ∂x1

=

∂ ∂Q (Tj − Rj ) − ∂x1 ∂xj

∂Tj ∂Q − j i ∂x ∂x =0 =

by (36).

We have now proved that the prolongations of an involutive tableau are involutive. One point of putting the argument here is that, whether one uses the language or not, the proof unavoidedly uses Spencer cohomology. Discussion. In Chapter VI, Theorem 2.1 we have proved that the prolongation of an involutive differential system (I, Ω) is again involutive. In case (I, Ω) is a linear Pfaffian system we may give an alternate proof based on Proposition 2.5 above as follows: To show that (I (1), Ω) is involutive on the manifold M (1) of integral elements of (I, Ω), we must show that (1) (1) i) for each point y ∈ M1 , the tableau Ay is involutive. ii) the integrability conditions for (I (1), Ω) are satisfied. Now (i) follows from (ii) in Proposition 2.5, and so we must establish (ii). We have seen in Chapter IV and will recall below that the integrability conditions of a linear Pfaffian differential system live in a family of quotient vector spaces. These equivalence classes were called the torsion of the linear Pfaffian system, and what we shall prove is that (38)

The torsion of (I, Ω) lives naturally in the family of vector spaces H 0,2 (Ax ), and

(39)

The torsion of (I (1), Ω) lives naturally in the family of vector spaces H 0,2 (A(1) ) ∼ = H 1,2 (Ax ) (cf. (26)). x

Since (I, Ω) is assumed to be involutive, its torsion vanishes. Then the torsion of (I (1), Ω) vanishes by the vanishing of cohomology assertion (i) in Proposition 2.5. This, at least in outline form, is the proof for linear Pfaffian systems of the fact that (I (1), Ω) is involutive if (I, Ω) is. We will now discuss how (38) and (39) are established, and for this we begin with a standard homological construction. Let A ⊂ W ⊗ V ∗ be a tableau and set B = A⊥ ⊂ W ∗ ⊗ V


283

so that we have an exact sequence of vector spaces 0 → A → W ⊗ V ∗ → B ∗ → 0. We define vector spaces B ∗(k) by 0 → A(k) → W ⊗ S k+1 V ∗ → B ∗(k) → 0

and set C

k,q

(B) =

B ∗(k−1) ⊗ Λq V ∗ 0

k≥1 k = 0.

Then we have an exact sequence of complexes 0 → C ·,·(A) → C ·,· → C ·,·(B) → 0 where C ·,· (A) = k,q C k,q (A), C ·,· = k≥0 (W ⊗S k+1 V ∗ ), and C ·,·(B) = k,q C k,q (B). Associated to this is a long exact cohomology sequence, and using (13) above this gives ∼ H p,q (B) −→ H p−1,q+1 (A), p ≥ 1 and q ≥ 1. In particular, we have (40)

(i) H 0,2 (A) ∼ = H 1,1 (B) 0,2 (1) ∼ (ii) H (A ) = H 1,2 (A) ∼ = H 2,1 (B).

To establish (38) we use equation (51) of Chapter IV to write the structure equations of (I, Ω) as dθa ≡ Aaεi π ε ∧ ωi +

1 a i c ω ∧ ωj 2 ij

mod {I}

where {I} is the algebraic ideal in Ω∗ M generated by the sections of I ⊂ T ∗ M . Under a change π ε → π ε + pεj ωj we have (41)

caij → caij +

1 a ε (A p − Aaεj pεi ). 2 εi j

Recalling that C 0,2 (A) = W ⊗ Λ2 V ∗ C 1,1 (A) = A ⊗ V ∗ H 0,2 (A) = W ⊗ Λ2 V ∗ /δ(A ⊗ V ∗ ) and that Ax is spanned by the Ax,ε = Aaεi (x)wa ⊗ vi∗ ,

284


may easily compute from (41) that the cocycle 1 a c (x)wa ⊗ vi∗ ∧ vj∗ ∈ C 0,2 (Ax ) 2 ij gives a class [c](x) in H 0,2 (Ax ) whose vanishing is necessary and sufficient for the existence of integral elements lying over x ∈ M (cf. Proposition 5.14 in Chapter IV for the complete argument here). If we write the structure equations in dual form as (cf. equation (83) of Chapter IV) a dθ ≡ πia ∧ ωi mod {I} Baλi πia ≡ Cjλ ωj mod {I} where the

B λ = Baλi wa∗ ⊗ vi

give a basis for the symbol relations in the annihilator Bx = A⊥ x, then the πia are determined modulo I up to a substitution πia → πia + paij ωj ,

paij = paji .

Under such a substitution Cjλ → Cjλ + Baλi paij , so that the equivalence class [Cjλ ] ∈ Bx∗ ⊗ V ∗ /δ(W ⊗ S 2 V ∗ ) = H 1,1 (Bx ) is well-defined. If we denote this equivalence class by [C](x), then it is easy to verify that: [c](x) = [C](x) under the isomorphism (i) in (40). We will now recall from equation (124) of Chapter IV the structure equations of the first prolongation (I (1) , Ω)

(42)

⎧ (i) dθa ≡ 0 mod {I (1) } ⎪ ⎪ ⎪ ⎨ (ii) dθa ≡ π a ∧ ωj mod {I (1) } where i ij a a ⎪ = πji (iii) πij ⎪ ⎪ ⎩ a λ k (iv) Baλi πij ≡ Cjk ω mod {I (1) }.

We will see that the equivalence class λ ] ∈ Bx∗(1) ⊗ V ∗ /δBx∗(2) [Cjk λ is well-defined, that the coboundary δ(Cjk ) = 0, and that the vanishing of the (1) resulting cohomology class [C ](x) ∈ H 2,1 (Bx ) is the necessary and sufficient condition for the existence of integral elements for (I (1), Ω) lying over x. Referring

§3. The Graded Module Associated to a Tableau; Koszul Homology

285

to (28) and (ii) in (39), we will then see that, in the involutive case, this vanishing is automatically satisfied. In summary, the involutivity of the tableau of (I, Ω) gives both the involutivity of the tableau and vanishing of the torsion of the prolonged system (I (1) , Ω). From equation (126) of Chapter IV we have λ j ω ∧ ωk ≡ 0 mod {I (1) }, Cjk λ which is just δ(Cjk ) = 0. Using (iv) in (31) it is now straightforward to show that (1) writing C as a coboundary is equivalent to being able to absorb the integrability a conditions into the πij , so that (iv) becomes a Baλi πij ≡0

mod {I (1) }.

In this way we have now established that (I (1), Ω) is involutive.

§3. The Graded Module Associated to a Tableau; Koszul Homology. We want to finish laying the basis for the proof of Theorem 2.4 above and for the Cartan–Kuranishi prolongation theorem in Chapter VI. The algebraic basis for both of these comes by studying a certain graded SV -module MA associated to a tableau A. A very interesting confluence occurs in that Cartan’s test for involution dualizes into the condition that MA admit a quasi-regular sequence, and such modules are standard fodder for the cannons of homological algebra. Before embarking on the formal discussion, we remark on why this should be so: The dual of ∂/∂xi : S q V ∗ → S q−1 V ∗ is the multiplication vi : S q−1 V → S q V. Thus, Cartan’s test in form of the surjectivity of the differentiation maps given by (6) above dualizes to the injectivity of suitable multiplication maps. Let A ⊂ W ⊗ V ∗ be a tableau and A=

(q) q≥0 A

⊂ W ⊗ S+ V ∗

the total prolongation of A. Dually, we set ⎧ ⊥ ∗ ⎪ ⎨ B =A ⊂W ⊗V Bq = A(q)⊥ ⊂ W ∗ ⊗ S q+1 V ⎪ ⎩ B = ⊕q≥0 Bq ⊂ W ∗ ⊗ S + V. For P ∈ W ⊗ S p+1 V ∗ and m ∈ W ∗ ⊕ S p V ∼ = (W ⊗ S p V ∗ )∗ we have (43)

0 / ∂P m, i = vi m, P ∂x

286


where vi ∈ V is a basis with dual basis xi ∈ V ∗ and , is the pairing between dual vector spaces. It then follows from (2) above that: B ⊂ W ∗ ⊗ S + V is the graded SV -submodule of

(44)

W ∗ ⊗ S + V generated by B ⊂ W ∗ ⊗ V . Here, W ∗ ⊗ SV is the obvious free SV -module and W ⊗ S + V the submodule corresponding to the maximal ideal in SV . We remark that since B = B0 , the ∗ grading on B is shifted by one from that induced by the natural grading on W ⊗SV . To rectify this we shall introduce the standard shift notation: if M = q Mq is a graded SV -module, then we define the new graded SV -module M [p] by (M [p] )l = Mp+l . With this notation the inclusion B[−1] → W ∗ ⊗ SV is a homogeneous SV -module mapping of degree zero. We define the graded SV module MA to be the quotient, so that we have 0 → B[−1] → W ∗ ⊗ SV → MA → 0.

(45) We note that

(46)

MA,q =

W∗

q=0

A(q−1)∗

q ≥ 1.

Definition 3.1. i) B is the symbol module associated to the tableau A, and ii) MA is the graded SV -module associated to the tableau A. To explain how Cartan’s test dualizes, we need to recall some essentially standard commutative algebra definitions. Let M = q≥0 Mq be a non-negatively graded SV -module. Definition 3.2. i) The element v ∈ V is quasi-regular if the kernel K in v

→M →0 0→K →M − of multiplication by v has no elements in positive degree; i.e., K + = (0). ii) The sequence v1 , . . . , vk of elements of V is quasi-regular for M if vj is quasiregular for M/(v1 , . . . , vj−1 )M for j = 1, . . . , k. iii) The module M is quasi-regular if there is a basis v1 , . . . , vn for V that is quasi-regular for M . The usual definitions of regular element, regular sequence, and regular module are the same as above but where multiplication has no kernel. The grading on the quotient module is the obvious one: (M/(v1 , . . . , vj−1)M )q = Mq /(v1 Mq−1 + · · · + vj−1 Mq−1 ). The point of all this is the following:


287

Proposition 3.3. The tableau A ⊂ W ⊗ V ∗ is involutive if, and only if, the associated graded module MA is quasi-regular. Proof. By Proposition 1.2 and Definition 1.3 above, the involutivity of A is equivalent to the existence of a basis x1 , . . . , xn ∈ V ∗ such that each mapping (q+1)

∂/∂xi : Ai−1

(47)

(q)

→ Ai−1

is surjective for all q ≥ 0. Recalling that by definition Ai−1 = {P ∈ A :

∂P ∂P = ··· = = 0} ∂x1 ∂xi−1

it follows from (43) and (44) that (q)⊥

Ai−1 = (v1 , . . . , vi−1 ) · MA,q ∩ ∩ A(q)∗ ∼ M = A,q+1 , and therefore (q) (q)⊥ (Ai−1 )∗ ∼ = A(q)∗ /Ai−1

∼ = MA,q+1 /(v1 , . . . , vi−1 ) · MA,q . The surjectivity of (47) for q ≥ 0 is equivalent to the injectivity of the dual mapping (47∗ )

vi : MA,q+1 /(v1 , . . . , vi−1 ) · MA,q → MA,q+2 /(v1 , . . . , vi−1 ) · MA,q+1

for q ≥ 0. This implies the proposition.

It is an remarkable coincidence of previously unrelated historical terminology that the condition that the constant coefficient Pfaffian differential system associated to the tableau A have a regular integral flag is equivalent to MA being quasi-regular in the sense of commutative algebra. We shall now quickly extend this discussion to a higher order tableau. Let A ⊂ W ⊗ S p+1 V ∗ be a tableau of order p with prolongations A(q) ⊂ W ⊗ S p+q+1 V ∗ . We set B = A⊥ ⊂ W ∗ ⊗ S p+1 V Bq = A(q)⊥ ⊂ W ∗ ⊗ S p+q+1 V B = q≥0 Bq , B0 = B. Using the shift notation we define the graded SV -module MA associated to A by the exact SV -module sequence (48)

0 → B[−1] → (W ∗ ⊗ SV )[p] → MA → 0.

288


Then

MA,q =

W ∗ ⊗ Sp V A(q)∗

q=0 q ≥ 1.

The reason for the choice of MA,0 will be discussed below. Remark on the obvious but interesting point that (W ∗ ⊗ SV )[p] = (W ∗ ⊗ S p V ) ⊕ (W ∗ ⊗ S p+1 V ) ⊕ . . . is not a free SV -module when p ≥ 1. This will be further discussed below. For now the important observation is that the statement and proof of Proposition 3.3 carry over verbatim to a higher order tableau. It is a well-known result in commutative algebra that the condition for a graded SV -module to admit a regular sequence is expressed by the vanishing of suitable Koszul homology groups. Moreover, the same proof works for quasi-regular sequences. We shall now explain this. Let M = q Mq be a graded SV -module with module mappings S p V ⊗ Mq → Mp+q . Set Cp,q = Mp ⊗C Λq V and define a boundary operator ∂ : Cp,q → Cp+1,q−1

(49) by the formula (50)

∂(m ⊗ vi1 ∧ · · · ∧ viq ) =

(−1)α+1 viα · m ⊗ vi1 ∧ · · · ∧ vîα ∧ · · · ∧ viq . α

It is clear that ∂ 2 = 0, and we set (51)

Hp,q (M ) = ker{∂ : Cp,q → Cp+1,q−1 }/∂Cp−1,q+1 .

Definition 3.4. Hp,q (M ) are the Koszul homology groups of the graded SV -module M. The Koszul homology groups will be used in this and some of the following sections. For the moment the following is the relevant property: Proposition 3.5. The following are equivalent: i) H+,n (M ) = · · · = H+,n−q (M ) = 0; ii) there is a quasi-regular sequence of length q, say v1 , . . . , vq , for M ; iii) every generic sequence of length q is quasi-regular for M . We denote by S + V the maximal ideal in SV and shall begin by establishing the


289

Lemma 3.6. The following are equivalent: i) H+,n (M ) = 0 ii) for m ∈ M , S + V · m = 0 ⇒ m ∈ M0 iii) there exists a v ∈ V such that v · m = 0 ⇒ m ∈ M0 iv) for generic v ∈ V , v · m = 0 ⇒ m ∈ M0 . Proof. We begin by showing the equivalence of i) and ii). The top end of the complex that computes Koszul homology is ∂

→ M ⊗ Λn−1V, 0 → H·,n (M ) → M ⊗ Λn V − where by the boundary formula (50) ∂(m ⊗ v1 ∧ · · · ∧ vn ) =

n

(−1)α+1 vα · m ⊗ v1 ∧ · · · ∧ vˆα ∧ · · · ∧ vn .

α=1

From this the equivalence of i) and ii) is clear. The equivalence of ii)–iv) is also pretty clear. What is obvious is that iv) ⇒ iii) ⇒ ii), and so we must prove that ii) ⇒ iv). Let J0 ⊂ M0 be a vector space complement to {M ∈ M0 : v · m = 0 for all v ∈ V } and set M = J0 ⊕ M + . Then M is a graded SV -module with the property: (52)

v · m = 0 for all v ∈ V ⇔ m = 0.

Referring to §1, n0 1 of Bourbaki [1961], with the terminology employed there we have that SV ∈ / Ass(M ). Moreover, by Corollary 2 in §1, n0 1 of Bourbaki [1961], the condition that the multiplication P : M → M by P ∈ SV be injective is that P ∈ / I for any prime ideal I ∈ Ass(M ). By the 0 corollary to Theorem 2 in §1, n 4 of Bourbaki [1961], the set Ass(M ) is a finite set I1 , . . . , Ik of proper prime ideals. Then SV \(I1 ∪ · · · ∪ Ik ) consists of the elements P such that multiplication by P is injective. Each Ij ∩ V is a proper linear subspace, and V \((I1 ∩ V ) ∪ · · · ∪ (Ik ∩ V )) is the open dense set of elements that are generic in the sense of (iv). We next need the following trivial but basic Lemma 3.7. Multiplication by v ∈ V induces the zero map v

Hp,q (M ) − → Hp+1,q (M ). Proof. If ϕ ∈ Mp ⊗ Λq V is a cycle, then ϕ ∧ v ∈ Mp ⊗ Λq+1 V and by (50) (53) This implies the lemma.

∂(ϕ ∧ v) = (∂ϕ) ∧ v + (−1)q v · ϕ.

290


Proof of Proposition 3.5. If v is quasi-regular then we have exact sequences of SV -modules v 0 → J0 → M − → vM → 0, J0 ⊂ M0 , 0 → vM → M → M/vM → 0, where in the first sequence v is a module map homogeneous of degree one, while all other maps are homogeneous of degree zero. By standard reasoning these give long exact homology sequences v ∂ → Hp,q (J0 ) → Hp,q (M ) − → Hp+1,q (vM ) − → Hp+1,q−1 (J0 ) → → Hp+1,q (vM ) → Hp+1,q (M ) → Hp+1,q (M/vM ) → Hp+2,q−1 (vM ) → . Using Lemma 3.7 and the fact that Hp+1,q−1 (J0 ) = 0 for p ≥ 0, this pair of sequences combines to give 0 → Hp+1,q (M ) → Hp+1,q (M/vM ) → Hp+1,q−1 (M ) → 0 for p ≥ 0. Now take q = n and use Lemma 3.6 to obtain (54)

Hp+1,n (M/vM ) ∼ = Hp+1,n−1(M ),

p ≥ 0.

If there is a v˜ ∈ V that is quasi-regular for M/vM , then we conclude from Lemma 3.6 that (55)

Hp+1,n−1 (M ) = 0,

p ≥ 0.

Conversely, if (55) holds then by (54) and the lemma we may find v˜ ∈ V that is quasi-regular for M/vM . Continuing in this way with an obvious descending induction gives the proposition. Corollary 3.8. The tableau A (of any order) p is involutive if, and only if, Hp,q (MA ) = 0,

for p ≥ 1, q ≥ 0.

Completion of the Proof of Theorem 2.4. We must show that the vanishing of suitable cohomology implies that the tableau A is involutive. By (43) and the definition of MA , the complexes of vector spaces δ

· · · → A(p) ⊗ Λq V ∗ − → A(p−1) ⊗ Λq+1 V ∗ → . . . · · · ← MA,p ⊗ Λq V ← MA,p−1 ⊗ Λq+1 V ← . . . and mutually dual. It follows that Spencer cohomology is dual to Koszul homology, i.e.,, (56)

H p,q (A) Hp,q (MA )∗ .

Under this duality, the above corollary translates into the implication H p,q (A) = 0 for p ≥ 1, q ≥ 0 ⇒ Ainvolutive. There is something missing here in the file that is in the book!


291

In order to establish the prolongation theorem in Chapter VI we need to prove the Proposition 3.9. Let A be a tableau. Then there is a k0 such that the prolongations A(k) are involutive for k ≥ k0 . We will actually prove a stronger result. To state it, we recall that there is a function, the Hilbert function PA (q) of the graded module MA , such that dim A(q) = PA(q) for all q. The result we shall actually need is given by the Proposition 3.10. Let A ⊂ W ⊗ V ∗ be a tableau and PA(q) the Hilbert function of the graded module MA . Then there is a k0 depending on dim W and PA(q) such that A(k) is involutive for k ≥ k0 . We will now prove Proposition 3.9. Then we shall prove Proposition 3.10 at the end of §5 after we have discussed localization. By Theorem 2.4, it will suffice to show that there is a k0 such that all (57)

H p,q (A(k) ) = 0,

p ≥ 1, q ≥ 0 and k ≥ k0 .

By the definition of Spencer cohomology we have (cf. (26) above) (58)

H p,q (A(k) ) ∼ = H p+k,q (A),

p ≥ 1 and q ≥ 0.

Combining (56)–(58) we see that Proposition 3.9 follows from (and in fact is equivalent to) the Proposition 3.9 . Let MA be the graded module associated to a tableau. Then there exists a p0 such that Hp,q (MA ) = 0 for p ≥ p0 , q ≥ 0. In fact, this proposition is valid for any finitely generated graded SV -module M , and it is this more general result that we shall prove. For this we set Cp,q = Mp ⊗ Λq V Cq = p Cp,q , and define a graded SV -module structure on Cq by v · (m ⊗ vi1 ∧ · · · ∧ viq ) = (v · m) ⊗ vi1 ∧ · · · ∧ viq where v, vi ∈ V and m ∈ M . The boundary mappings (49) induce (59)

∂ : Cq → Cq−1 ,

and it is immediate from (50) that (59) is a mapping of graded SV -modules, homogeneous of degree one. As a consequence, assuming that M· is finitely generated the following are all finitely generated SV -modules ker ∂, image ∂, Hq (M ) = p Hp,q (M ). Moreover, and this is the crucial point, by Lemma 3.7 above the maximal ideal S + V acts trivially on the SV -module Hq (M ). From this we infer that Hq (M ) is a finitely generated SV /S + V = C module, i.e., it is a finite dimensional vector space.

292


In the next section we shall discuss the interpretation of Koszul homology Hp,q (M ) as it pertains to resolutions of M by free modules. It will turn out that Proposition 3.9 is really just the statement that a sub-module of a finitely generated SV ∗ module is itself finitely generated as this latter statement pertains to the relations among the generators of M , and then the relations among the relations, and so forth, i.e., it is equivalent to the statement that M has a finite resolution by finitely generated free modules.

§4. The Canonical Resolution of an Involutive Module. In the preceeding section we defined the graded module MA associated to a tableau, and then we used more or less standard commutative algebra to relate the existence of quasi-regular sequences for MA to the vanishing of Koszul homology. There the motivation was to complete the proof of Theorem 2.4, among other things providing the conceptual basis for the result that the prolongations of an involutive tableau are involutive. Another use of Koszul homology is in the construction of resolutions of graded modules, and we would now like to pursue the implications of this for the theory of differential systems. Before doing this we will try to give some motivation by the following Example 4.1. We identify W ∗ ⊗ SV with constant coefficient linear differential operators on W ∗ ⊗ SV ∗ , so that by definition (wa∗ ⊗ vI )(PJb wb ⊗ xJ ) =

∂ (P a xJ ). ∂xI J

As discussed in Chapter IV, giving a tableau A ⊂ W ⊗ V ∗ is equivalent to giving the linear homogeneous constant coefficient P.D.E. system (60)

Dλ u(x) = 0

where Dλ = Baλi wa∗ ∂/∂xi u = wa ua (x) Dλ u = Baλi and where the

∂ua (x) , ∂xi

B λ = Baλi wa∗ ⊗ vi

give a basis for B = A⊥ ⊂ W ∗ ⊗ V . The symbol module B corresponds to the algebra of constant coefficient differential operators on W ⊗ SV ∗ generated by the operators Dλ . In particular, the solutions to (60) satisfy Du = 0 for all D ∈ B.

§4. The Canonical Resolution of an Involutive Module

293

More generally we consider the inhomogeneous linear P.D.E. system Dλ u(x) = f λ (x).

(61) Any relation

eλ B λ = 0,

eλ ∈ SV

on the generators of the symbol module gives an integrability condition Eλ f λ (x) = 0 on the f λ , where Eλ ∈ C[∂/∂x1 , . . . , ∂/∂xn ] corresponds to eλ ∈ C[v1 , . . . , vn ] when we set vi = ∂/∂xi . In this way, not only the generators but also the relations of the symbol module enter into the theory. Once we agree to study the relations of B, we may as well go ahead and study entire resolutions. In fact, returning to the general discussion, it is well-known that in some ways the most important properties of a finitely generated SV -module M are given by its generators and relations, and then the generators and relations of its relations, etc. In brief, one wants to find resolutions of M by free SV -modules E = E ⊗ SV , where E is a finite-dimensional vector space. It is also well-known that any M has an essentially canonical minimal such resolution where the vector spaces E are appropriate Koszul homology groups of M . It has been proved above that the involutivity of A is equivalent to the vanishing of certain of the Koszul homology groups of MA , which then turns out to be equivalent to the property that the canonical resolution be especially simple (this was conjectured by Guillemin and Sternberg and proved by Serre; see Guillemin and Sternberg [1964]). Here we shall recall some of the definitions and elementary facts, together with a derivation of the canonical resolution of an involutive module. We set S = SV and consider an arbitrary finitely generated, graded S-module M = k∈Z Mk . For simplicity of exposition we shall assume that M is nonnegatively graded in the sense that Mk = 0 for k < 0. The action of S on M is given by vector space mappings Sj ⊗ Mk → Mj+k satisfying the customary conditions. We let S + = k≥1 S k V be the maximal ideal of S; then M/S + M is a finite dimensional vector space whose dimension equals the minimal number of generators of M as an S-module. Associated to M are its Koszul homology groups Hk,q (M ), whose definition was given in the preceeding section. Here we shall give a number of remarks concerning these groups (cf. Green [1989a] for a general discussion of Koszul homology and its relationship to algebraic geometry): (62)

Hk,0 (M ) ∼ = Mk /V · Mk−1 ∼ = {new generators of M in degree k}.

In particular (63)

(64)

M is generated in degree zero

⇔ Hk,0 (M ) = 0 for all k > 0.

Hk,n (M ) ∼ = AnnV (Mk ) = {v ∈ V | v · Mk = 0}.

294


Of importance is the shift mapping, which we recall associates to a graded module M a new graded module M [p] defined by (M [p] )q = Mp+q .

(65)

It is then clear from the definitions that (66)

Hk,q (M [p] ) ∼ = Hk+p,q (M ) for q ≥ 0.

An exact sequence of graded S-modules (67)

0 → M → M → M → 0

where all maps are homogeneous of degree zero, gives a long exact homology sequence (68)

→ Hk−1,q+1 (M ) − → Hk,q (M ) → Hk,q (M ) → Hk,q (M ) ∂

− → Hk+1,q−1(M ) → . . . . ∂

In case the module maps in (67) are homogeneous of degrees other than zero, then we still have an exact sequence (68) with a shift in indices given by using the shift mapping to make the maps homogeneous of degree zero. An example when this shift occurs is given by the exact S-module sequence 0 → B → W ∗ ⊗ SV → MA → 0 where the first map is homogeneous of degree +1. Here, A is an ordinary tableau, B is its symbol module, and MA is the associated graded SV -module. The sequence (45) above 0 → B[−1] → W ∗ ⊗ S → MA → 0 has all maps homogeneous of degree zero, and (68) together with (66) gives (70) ∂ · · · → Hk,q (W ∗ ⊗ SV ) → Hk,q (MA ) − → Hk,q−1(B) → Hk+1,q−1 (W ∗ ⊗ SV ) → . . . (71)

M is a free module generated in degree zero (i.e., M ∼ = M0 ⊗ SV ) if, and only if,

(72)

Hk,q (M ) = 0 for (k, q) = (0, 0).

This property will be proved when we establish Proposition 4.3 below. We recall our assumption that all modules are non-negatively graded; thus M = q≥0 Mq . Definition 4.2. i) A graded S-module M is involutive if (73)

Hk,q (M ) = 0 for k ≥ 1 and all q ≥ 0


ii) A graded S-module M =

k≥k0

295

Mk is k0 -involutive if

Hk,q (M ) = 0 for k = k0 and all q ≥ 1. In the latter case it follows that ˜ = M [k0 ] M is involutive in the usual sense. We may reformulate the discussion in the preceeding section by the statement: (74)

The tableau A is involutive if, and only if, the associated graded module MA is involutive.

From (70) and (72) we have the following: (75)

The tableau A is involutive if, and only if, the symbol module B, is involutive.

By applying the standard construction of a minimal free resolution of a graded S-module in terms of its Koszul homology groups, we shall derive the following result due to Serre (see Guillemin and Sternberg [1964]): Proposition 4.3. Let M be an involutive SV -module. Then there is a free resolution (76)

ϕn

ϕ1

ϕ0

0 → En −−→ En−1 → · · · → E1 −→ E0 −→ M → 0

where deg ϕ0 = 0, deg ϕi = 1, for i ≥ 1 and Ei = H0,i (M ) ⊗ SV . The converse is also true. We shall call (76) the canonical resolution of an involutive module (it is canonical in the sense that it is minimal in the sense explained below). Before deriving it we shall make a few remarks. (77)

An involutive module is generated in degree zero, i.e., M = S · M0 .

This follows from the definition together with (63). Next we consider an exact sequence 0 → M → M → M → 0 of graded SV -modules and degree zero maps. (78)

If M, M are involutive and M0 → M0 is an isomorphism, then M is 1-involutive. Moreover, there are canonical short exact sequences

296

(79)


0 → H0,q (M ) → H0,q (M ) − → H1,q−1 (M ) → 0. ∂

This is clear from our definitions and (68). For 1-involutive modules M we shall use the notation ˜ = M [1] . M ˜ is involutive and (79) is Then under the hypotheses of (19), M (80)

˜ ) → 0. 0 → H0,q (M ) → H0,q (M ) → H0,q−1(M

We may now easily give the construction of the canonical resolution of an involutive module. By (62) and (63) H0,0 (M ) ∼ = M0 , and M is generated in degree zero. Setting E0 = H0,0(M ) ⊗ SV , this gives a short exact sequence 0 → N → E0 → M → 0 to which (78) and (79) apply. Thus N˜ is involutive, and by (71) and (80) ˜ ), H0,q (M ) ∼ = H0,q−1 (N

q ≥ 1.

˜ , and continue. We now repeat the construction replacing M by N Note that if M is free and H0,1 (M ) = 0 in (77), then M = 0 and so M ∼ = M is free. In other words, if N is involutive and H0,1 (N ) = 0, then N is free. This implies that the resolution process terminates after at most n steps. We have now established the constructive half of Proposition 4.3; since we will not use the converse the proof of this will not be given (in any case it is standard). Discussion. We shall, without proofs, put Proposition 4.3 in a general context. For this we let (i)

ϕm

ϕ1

ϕ0

0 → Em −−→ En−1 → · · · → E1 −→ E0 −→ M,

deg ϕi = 0,

be a resolution of a finitely generated graded S-module M by free modules Ep = q (Bp,q ⊗ S [−q] ), where the Bp,q are finite dimensional vector spaces whose dimensions bp,q give the number of S [−q] ’s occurring in Ep . The maps ϕi : Ei → Ei−1 are normalized to be of degree zero and are given by a matrix whose entries break into blocks corresponding to maps (ii)

ϕi,k,l : Bi,k ⊗ S [−k] → Bi−1,l ⊗ S [−l] .

Each such block is clearly given by a bi,k × bi−1,l matrix whose individual entries are homogeneous polynomials of degree k − l. We shall say that the resolution (i) is minimal in case the entries in each ϕi,k,l have strictly positive degree.


297

To explain this, it is clear that a non-zero block (ii) where k = l gives a redundancy in (i) in the sense that in terms of suitable bases for Bi,k and Bi−1,k the matrix of ϕi,k,l will have the form Iα 0 , 0 0 where Iα is the α × α identity matrix and where the Iα piece induces isomorphism between free sub-modules of Ei and Ei−1 which may then be deleted from (i). It follows that every M has a minimal free resolution (i), and this resolution may be seen to be essentially unique. In fact, a general result given for example in Green [1989a] is that (iii)

Hp,q (M ) ∼ = Bq,q+p .

We will briefly discuss this result. We recall that M = q≥0 Mq is assumed to be non-negatively graded. Then (iii) implies that (iv)

Bp,q = 0 for q < p.

In other words, the generators of M (which correspond to the B0,q ) are in nonnegative degrees, the relations among the generators (which correspond to the B1,q ) are themselves generated in degrees at least one, and so forth. Using (iv) we may write (v) Eq = p≥0 (Cp,q ⊗ S [−q−p] ) where by (iii) (vi)

Cp,q ∼ = Hp,q (M ).

Now the simplest modules, the free modules, are those for which Cp,q = 0,

q = 0.

By (vi) these are characterized by Hp,q (M ) = 0,

q = 0.

The next simplest modules, at least from our point of view, are those for which (vii)

Cp,q = 0,

p = 0.

By (vi) and the Definition 4.2 these are just the involutive modules, and clearly (v) above is equivalent to Proposition 4.3. Generalizing slightly the above, we may see from (vi) and the definition that a finitely generated module M is k0 -involutive if, and only if, its minimal resolution (i) has Ep = Bp ⊗ S [−k0 −p] where (viii) Bp = Hk0,p (M )

298


Example 4.4. The truncation Sk0 of S is defined by q S V q ≥ k0 (Sk0 )q = 0 q < k0 . It is easy to verify that Hp,q (Sk0 ) = 0 for p = k0 . The minimal resolution of M may be constructed using Young symmetrizers (cf. Green [1989a]). When k0 = 1 the truncation is the maximal ideal S + V . Resolving S + V is equivalent to resolving SV /S + V = C, and this is provided by the standard Koszul resolution ∂

∂

∂

... − → S ⊗ Λ2 V − → S ⊗V − → S → C → 0,

deg ∂ = 1,

obtained by dualizing the polynomial de Rham complex 0 → C → SV ∗ − → SV ∗ ⊗ V ∗ − → SV ∗ ⊗ Λ2 V ∗ − → ... d

d

d

and applying Proposition 2.1 above. The inhomogeneous P.D.E. system whose symbol module is Sk0 is ∂ k0 u(x) = fI (x), ∂xI

|I| = k0 .

The involutivity of this system implies in particular that the compatibility conditions for this system are all of 1st order (see the discussion below). When k0 = 1 the system is ∂u(x) = fi (x), ∂xi and the compatibility conditions are obviously ∂fi (x) ∂fj (x) = . j ∂x ∂xi The remainder of this section will be a series of remarks, some of which are of interest in themselves and some of which are for later use. (81)

A piece of the canonical resolution is (where we set S m V = 0 for m < 0)

(82)

0 → H0,n (M ) ⊗ S m V → H0,n−1(M ) ⊗ S m+1 V → . . . → H0,1 (M ) ⊗ S m+n−1 V → H0,0(M ) ⊗ S m+n V → Mm+n → 0.

Of importance below will be the explicit description of the maps ϕi : H0,i(M ) → H0,i−1(M ) ⊗ V occurring in the resolution (76). For this we now denote the Koszul boundary by ∂V : Λq V → Λq−1 V ⊗ V where ∂V (vi1 ∧ · · · ∧ viq ) =

α+1 vi1 α (−1)

∧ · · · ∧ vîα ∧ · · · ∧ viq ⊗ viα . Recalling that


299

H0,i (M ) = ker{∂ : M0 ⊗ Λi V → M1 ⊗ Λi−1 V }1 ∩ M0 ⊗ Λi V we infer the commutative diagram ϕi

H0,i (M ) −→ H0,i−1 (M ) ⊗ V ∩ ∩ 1⊗∂V M0 ⊗ Λi−1 V ⊗ V. M0 ⊗ Λi V −−−→ For ξ ∈ M0 ⊗ Λi V , the cycle condition ∂ξ = 0 is that the composite map 1⊗∂

V M0 ⊗ Λi V −−−→ M0 ⊗ Λi−1 V ⊗ V → M1 ⊗ Λi−1 V

applied to ξ be zero, i.e., identifying Λi−1 V ⊗ V with V ⊗ Λi−1 V , (83)

(1 ⊗ ∂V )(ξ) ∈ (degree one relations for M ) ⊗ Λi−1 V ∩ (M0 ⊗ V ) ⊗ Λi−1 V.

This implies that (1 ⊗∂V )(ξ) lies in the subspace H0,i−1(M )⊗V of M0 ⊗Λi−1 V ⊗V and that ϕi is induced by 1 ⊗ ∂V . (84)

In case M = MA for involutive tableau A ⊂ W ⊗ V ∗ with symbol relations B ⊂ W ∗ ⊗ V , the first few pieces of the canonical resolution (82) are (using (83)) 0 → H0,0(MA ) ∼ = (MA )0 → 0; W∗ W∗ 0 → H0,1(MA ) → H0,0 (MA ) ⊗ V → (MA )1 → 0 0→ B → W∗ ⊗ V → A∗ → 0,

and (more interestingly) 0→H0,2 (MA )→H0,1 (MA )⊗V →H0,0 (MA )⊗S 2 V →(MA )2 →0

(85)

∩

W ∗ ⊗Λ2 V →

B⊗V

→

W ∗ ⊗S 2 V

→ A(1)∗ →0.

Referring to Example 4.1 above, we let Dλ = Baλi wa∗ ⊗ vi ∈ W ∗ ⊗ V 1 For

i = 1 this just gives that H0,1 (M ) ∼ = ker(M0 ⊗ V → M1 ) ∼ = degree one relations for M .

300


be a basis for B. Then by (85), elements of H0,2 (MA ) are of the form q=

1 ij ∗ q w ⊗ ∂/∂xi ∧ ∂/∂xj ∈ W ∗ ⊗ Λ2 V 2 a a

where qaij + qaji = 0 and qaij = mjλ Baλi . Since H0,2 (MA ) ∼ = H0,1(B) ∼ =

relations among the generators Dλ of B

,

it follows that: (86)

All relations among the Dλ = Baλi wa∗ ⊗ ∂/∂xi are generated by linear relations of the form mjλ ∂/∂xj Dλ = 0.

(Remarkably, this fact is found in Cartan [1953].)2 Referring now to the inhomogeneous, constant coefficient linear P.D.E. system (cf. (61) above) Dλ u(x) = f λ (x), we see that the compatibility conditions on the f λ (x) are generated by the 1st order equations ∂f λ (x) mjλ = 0. ∂xi (87)

The dual of the graded piece (82) of the canonical resolution of the involutive module MA associated to an involutive tableau A is (setting q = m + n)

(88)

0 → A(q) → H 0,0 (A) ⊗ S q+1 V ∗ → H 0,1 (A) ⊗ S q V ∗ → H 0,2 (A) ⊗ S q−1 V ∗ → . . .

Although we do not need it for our work here we remark that (88) is essentially the symbol sequence of the Spencer complex associated to an involutive linear P.D.E. system (cf. Chapter X). to (85), the kernel H0,1 (B) of B ⊗ V → W ∗ ⊗ Λ2 V is identified with all mjλ D λ ⊗ ∈ B ⊗V ⊂ W ∗ ⊗V ⊗V that lie in W ∗ ⊗Λ2 V ; this is just the condition mjλ Baλi +miλ Baλi = 0, j which is clearly equivalent to mλ ∂/∂xj D λ = 0. The reference to Cartan is page 1045 in the 1984 edition of his collected works. 2 Referring

∂/∂xj


301

For later purposes it is useful to make explicit the maps in (88). First note that, by the definitions, H 0,i (A) = W ⊗ Λi V ∗ /δ(A ⊗ Λi−1 V ∗ ).

(89)

We claim that δ induces W ⊗ Λi V ∗ W ⊗ Λi+1 V ∗ k ∗ δ (90) → ⊗S V − ⊗ S k−1 V ∗ , δ(A ⊗ Λi−1 V ∗ ) δ(A ⊗ Λi V ∗ ) which is the map induced by exterior differentiation δ : W ⊗ S k V ∗ ⊗ Λi V ∗ → W ⊗ S k−1 V ∗ ⊗ Λi+1 V ∗ (thinking of the S k V ∗ as polynomial functions). Proof. Let ψ ∈ A ⊗ Λi−1 V ∗ and P ∈ S k V ∗ . Then, from the definition of δ, since δψ ∈ W ⊗ Λi V ∗ are constant coefficient differential forms δ(δψ ⊗ P ) =

(−1)i δψ ∧ dxj ⊗

i

=

∂P ∂xj

(−1)i δ(ψ ∧ dxj ) ⊗

j

∂P ∈ (δ(A ⊗ Λi V ∗ )) ⊗ S k−1 V ∗ . ∂xj

From this and (83) it follows that: (91) (92)

The maps in the long exact sequence (88) are induced by δ in (90). Finally, we want to discuss the “1st derived part” of an involutive module M . Recall from (64) that Hk,n(M ) = 0 k ≥ 1 H0,n(M ) = {m ∈ M0 : v · m = 0 for all v ∈ V } = W0∗

where the last equation is a definition of W0∗ . Setting N = M/W0∗ we have an exact sequence of graded SV -modules and degree zero maps (93)

0 → W0∗ → M → N → 0

where N has the following properties (both of which come from the exact Koszul homology sequence of (93)): (i) N is involutive (ii) H0,n(N ) = 0. Moreover, the sequence (93) gives (i) 0 → W0∗ → W ∗ → N0 → 0 (94) ∼ (ii) Mk −→ Nk (k ≥ 1).

(degree zero)

302


From (i) we have an intrinsic subspace W1 = W0∗⊥ ⊂ W. Now suppose that M = MA for an involutive tableau A. From (ii) in (94) in the case k = 1 we have (95)

A ⊂ W1 ⊗ V ∗ ⊂ W ⊗ V ∗ .

Definition 4.5. i) We shall say that two tableau Ai ⊂ Wi ⊗V ∗ , i = 1, 2 are equivalent if there is a vector space W and inclusions Wi ⊂ W such that A1 = A2 as subspaces of W ⊗ V ∗ . ii) We shall say that two symbol mappings σi : Wi ⊗ V ∗ → Ui are equivalent in case the tableau Ai = ker σi are equivalent. To see what this means, suppose that W2 = W and we have the situation (95). Choose a direct sum decomposition W = W0 ⊕ W1 and basis for W compatible with this composition. Then the tableau matrix πia for A has the form 0 πiρ where πiρ is the tableau matrix for A ⊂ W1 ⊗ V ∗ . In the language of linear Pfaffian differential systems, assuming involutivity we easily see that the 0-block (corresponding to W0∗ ⊂ W ∗ ) gives the 1st derived subsystem of our Pfaffian differential system. Thus: For an involutive, linear Pfaffian differential system the 1st derived system corresponds to the subspaces H0,n (MA ) ⊆ W . Below we shall introduce a refinement of the characteristic variety called the characteristic sheaf, and in the next section shall prove that (96)

the characteristic sheaf of an involutive linear Pfaffian differential system uniquely determines the symbol mapping up to equivalence.

There are examples to show that the corresponding statement for the characteristic variety is false.

§5. Localization; the Proofs of Theorem 3.2 and Proposition 3.8

303

§5. Localization; the Proofs of Theorem 3.2 and Proposition 3.8. We continue with our purely algebraic discussion centered around the algebraic properties of an involutive tableau A ⊂ W ⊗ V ∗ . It is well-known that a major advance in commutative algebra occurred by localizing, or by what is essentially equivalent, by the use of sheaf theory. It is also well-known that a major advance in linear P.D.E. theory occurred by microlocalizing in the cotangent bundle. It is therefore reasonable to adapt these two techniques to the theory of linear Pfaffian differential systems in the expectation that they may prove useful in the study of geometric problems. That is what we shall do in this section, which will be broken into a number of discussions. In order to simplify notation, we make the convention for this section: Unless mentioned otherwise, all vector spaces will be assumed to be complex. Preliminaries. For a complex space V we set P = P V ∗ with homogeneous vector k coordinate ring S = SV = k≥0 S V and with O denoting the structure sheaf on P . We will use the well known F.A.C. “dictionary”, cf. Serre [1955] and Hartshorne [1977].

(97)

⎧ ⎫ ⎧ ⎫ ⎨ coherent sheaves of ⎬ ⎨ graded S-modules ⎬ O-modules and sheaf ↔ and S-module . ⎩ ⎭ ⎩ ⎭ maps over P maps

We remark that all modules are assumed to be of finite type. The italics around the word dictionary signify that the correspondence (97) is not a bijection. In the study of involutive, linear Pfaffian differential systems what will be lost are the graded modules associated to the 1st derived system part of the tableau; this will exactly correspond to the equivalence relation introduced in Definition 4.5 above. We will denote by O(q) the usual sheaf of “locally holomorphic homogeneous functions of degree q” on P . It is standard that H 0 (P, O(q)) = S q V (this holds for all q if we agree to set S q V = 0 for q < 0). Given a coherent sheaf F on P the corresponding graded S-module is

F =

q Fq 0

Fq = H (P, F (q)) where F (q) = F ⊗O O(q). The module structure S p V ⊗ Fq → Fp+q is obtained from the sheaf pairing O(p) ⊗O F (q) → F (p + q) by passing to global sections.

304


The inverse map graded S-module F → coherent sheaf F is obtained by localization. The following are some of its basic properties: i) If E = E ⊗C S is a free S-module, then the localization E is the trivial vector bundle with fibre E (in general we shall identify holomorphic vector bundles and locally free sheaves); ii) If the graded module F localizes to F , then the shift F [q] localizes to F (q); iii) exact module sequences go to exact sheaf sequences (but not quite conversely); and iv) Fq = H 0 (P, F (q)) for q ≥ q0 (F ). To construct the localization, we may use the fact that F has finite free resolution and use i). Remark that a module map ϕ:E→F between free modules E = E ⊗ S and F = F ⊗ S will by definition be homogeneous of some degree q ≥ 0; it is given by an element ϕ ∈ Hom(E, F ) ⊗ S q V , which we may think of as a global section of Hom(E(−q), F ). Thus ϕ localizes to ϕ : E(−q) → F , and in this way we know what is meant by the maps appearing in a free resolution. A basic general fact is that the dictionary (97) is exact modulo finite dimensional vector spaces. This is illustrated by the following Example 5.1. Let P1 , . . . , Pm ∈ S q V be homogeneous forms of degree q on P . There is then an exact sequence of graded S-modules (98)

m

ϕ

S− → S [q] → Q → 0

where the maps have degree zero and ϕ(Q1 , . . . , Qm ) = i Qi Pi , Q = coker ϕ

.

m S) is (a shift of) the homogeneous ideal {P1 , . . . , Pm} generThe image I = ϕ( ated by the Pr ’s. The sheaf sequence corresponding to (98) is (99)

m

ϕ

O− → O(q) → Q → 0

where Q is a coherent sheaf supported on the algebraic variety Z ⊂ P defined by the ideal I. In particular, if Z = ∅ then Q = (0) and (99) is m

ϕ

O− → O(q) → 0.

By the general theory (Theorem A in Serre [1955] to be specific), the induced mapping m 0 H (P, O(p)) → H 0 (P, O(p + q))


305

is surjective for p ≥ p0 ; this is a special case of Hilbert’s nullstellensatz. In this case Q = S/I is a finite dimensional vector space that is “lost” in the dictionary (97). The simplest special case is when Pi = ξi so that

Q = C[ξ1 , . . . , ξn ]/{ξ1 , . . . , ξn } ∼ = C.

This example also explains why the finiteness Theorem 3.12 in Chapter V should be true. Namely, consider the linear homogeneous P.D.E. system for one unknown function (100)

Pr (D)u(x) = 0,

r = 1, . . . , m

where Pr (D) ∈ C[∂/∂x1 , . . . , ∂/∂xn ] is the constant coefficient operator corresponding to Pr ∈ S q V . If the complex characteristic variety {[ξ] ∈ P : all Pr (ξ) = 0} of (100) is empty, then by the nullstellensatz we have that ξ α ∈ {P1 (ξ), . . . , Pm(ξ)} for |α| ≥ p0 + q; this gives Dα u(x) = 0 for |α| ≥ p0 + q. Thus u(x) is a polynomial of degree at most p0 + q, and therefore the solution space to (100) is finite dimensional. In fact, the solution space is naturally isomorphic to Q∗ (= dual vector space to Q). In the special case when P1 = ξ1 , . . . , Pn = ξn we obtain only the constant functions. It will come out of our discussion that this is the only involutive system (100) with empty characteristic variety. This is a special case of Corollary 3.11 to Theorem 3.6 in Chapter V. The Characteristic Sheaf. We consider a graded S-module M that has a presentation (101)

ϕ1

ϕ0

U ∗ ⊗ S −→ W ∗ ⊗ S −→ M → 0

where U, W are finite dimensional vector spaces and ϕi has degree i for i = 0, 1 (we may think of ϕ1 as a matrix whose entries are linear functions). The localization of (101) will be denoted by (102)

ϕ1

ϕ0

U ∗ (−1) −→ W ∗ −→ M → 0

where U ∗ , W ∗ are the trivial bundles with respective fibres U ∗ , W ∗ . Definition 5.2. We call M the characteristic sheaf of M and support(M) = Ξ the characteristic variety of M . It may be proved that both these definitions are independent of the particular presentation (101). To explain our motivation for this terminology, we let A ⊂ W ⊗ V ∗ be a tableau and set U = W ⊗ V ∗ /A ∼ = B∗

306


so that we have the symbol mapping σ : W ⊗ V ∗ → U. Let MA be the graded S-module associated to A as given by Definition 3.1 above. Definition 5.3. We call MA the characteristic sheaf of the tableau and ΞA = support(MA ) the characteristic variety of the tableau. In this special case (101) is σ∗

B ⊗ S −→ W ∗ ⊗ S → MA → 0, which is just the definition of MA , and the localization (102) is σ∗

B∗ (−1) −→ W ∗ → MA → 0. For each [ξ] ∈ P we define

σξ : W → U

by σξ (w) = σ(w ⊗ ξ). Clearly, σξ is defined only up to non-zero multiples. In intrinsic terms, if Lξ ⊂ V ∗ is the line corresponding to [ξ] ∈ P , then σξ = σ | W ⊗ Lξ . This induces the mapping which dualizes to

W → U ⊗ L∗ξ B ⊗ Lξ → W ∗ .

Now Lξ is the fibre of O(−1) at [ξ] so that the last mapping is σξ∗

∗ B(−1)[ξ] −→ W[ξ] ,

where the subscript denotes the fibres of the various bundles at [ξ] ∈ P . Our conclusion is that σξ∗ is the mapping σ ∗ localized at [ξ]. Consequently, if we denote by F[ξ] = F /m[ξ] · F the fibre of a coherent sheaf F at [ξ] where m[ξ] ⊂ O[ξ] is the maximal ideal, then (MA )[ξ] = (ker σξ )∗ . Moreover, support MA = ΞA is the characteristic variety of the tableau, defined set-theoretically as ΞA = {[ξ] ∈ P : ker σξ = 0}.


307

We may summarize by saying that the characteristic sheaf of a tableau contains not only the information of where the symbol fails to be injective, but by how much it fails to be injective. Involutive Sheaves. Returning to the general discussion we give the following Definition 5.4. An involutive sheaf is a coherent sheaf F that has a presentation ψk−1

ψk

(103)

0 → Fk (−k) −→ Fk (−(k − 1)) −−−→ · · · → ψ1

ψ0

F1 (−1) −→ F0 −→ F → 0

for some k ≤ n and where Fi = Fi ⊗C O is a trivial vector bundle. We shall give two remarks concerning this definition. The first is that an involutive module has been defined (cf. Definition 4.2 above) to be one whose Koszul homology groups have a certain vanishing property. Now associated to a coherent sheaf are both its cohomology groups H i (F (q)) = H i (P, F (q)) i and Koszul groups Kp,q (F ), the latter defined to be the middle homology of the 3-term complex H i (F (p − 1)) ⊗ Λq+1 V → H i (F (p)) ⊗ Λq V → H i (F (p + 1)) ⊗ Λq−1 V i (these are the usual Koszul groups for the graded S-module q H (F (q))). We may then equivalently define an involutive sheaf by the vanishing conditions i H (F (q)) = 0 for i = 0, q ≥ 0 0 (F ) = 0 Kp,q

for p ≥ 1, q ≥ 0.

For our purposes, however, it is more convenient to take the existence of a presentation (103) as our definition. Our second remark is that the presentation (103) is not unique. For example we consider the localized Koszul complex (104)

∂

∂

∂

∂

0 → Λn V ⊗ O(−n) − → ... − → Λ2 V ⊗ O(−2) − → V ⊗ O(−1) − →O→0

obtained by choosing a basis v1 , . . . , vn for V = H 0 (P, O(1)) and defining, as always, ∂(vi1 ∧ · · · ∧ viq ⊗ f) =

q

(−1)α+1 vi1 ∧ · · · ∧ vîα ∧ · · · ∧ viq ⊗ viα · f,

α=1

for 1 ≤ i1 , . . . , iq ≤ n and f ∈ O(−q). Although not strictly necessary for our purposes, it may be shown that: (105)

For F involutive, any two resolutions (103) differ by a direct sum of resolutions (104).

(106)

If we define (103) to be minimal in case ∼

H 0 (P, F0 ) −→ H 0 (P, F ) is an isomorphism, then any involutive F has a unique minimal resolution (103) with k ≤ n − 1.

308


We next observe that (107)

The characteristic sheaf of an involutive module M is an involutive sheaf.

In fact, the localization of the canonical resolution (cf. Proposition 4.3 above) ϕn−1

ϕn

ϕ1

ϕ0

0 → En −−→ En−1 −−−→ · · · → E1 −→ E0 −→ M → 0 of M gives a presentation ϕn−1

ϕn

ϕ0

ϕ1

0 → En (−n) −−→ En−1 (−(n − 1)) −−−→ · · · → E1 (−1) −→ E0 −→ M → 0 of the type (103). We now shall prove the converse: Proposition 5.5. An involutive sheaf is the characteristic sheaf of an involutive module. Proof. What we want to do is take global sections of (103) tensored with O(r) for each r ≥ 0. In order for this to work, we need (as always) a vanishing theorem. Lemma 5.6. Let F be an involutive sheaf. Then i) H i (P, F (q)) = 0 for i ≥ 1, q ≥ 0; ii) F is generated by its global sections; and iii) Fn∗ ∼ = ker{H 0 (P, F0 ) → H 0 (P, F )} (this is understood to be zero if k < n in (103)). Proof. Although one may prove this result directly by induction on k, the argument is clearer if we use spectral sequences. For each r we obtain from (103) the long exact sheaf sequence (108)

0 → Fk (−k + r) → Fk−1 (−(k − 1) + r) → · · · → F1 (−1 + r) → F0 (r) → F (r) → 0

We may view (108) as a complex of sheaves whose cohomology sheaves are zero. Associated to any complex of sheaves are two spectal sequences both having the same abutment. Since the cohomology sheaves of (108) are trivial, one spectral sequence has its E1 terms equal to zero and therefore the abutment is trivial. If we label the terms in the complex (108) with Fk (−k + r) corresponding to the index 0 and F (r) to the index k + 1, then the other spectral sequence has

E1p,q = H q (P, Fk−p(−(k − p) + r)) E1k+1,q

= H (P, F (r)) q

0≤p≤k

.

We shall use the well-known fact that ⎧ ⎪ ⎨ 0 1 ≤ s ≤ n − 2 and all q q (109) H (P, O(s)) = 0 q = n − 1 and s ≥ −(n − 1) . ⎪ ⎩ C q = n − 1 and s = −n


309

For r ≥ 0 the only possible non-zero terms in the spectral sequence are therefore E1p,0 = H 0 (P, Fk−p(−(k − p) + r)) ∼ = Fk−p ⊗ S −k+p+r V E1k+1,0 = H 0 (P, F (r)) E10,n−1 = H n−1 (P, Fn (−n)) ∼ = Fn∗ in case k = n, r = 0 (here we recall our convention that S t V = (0) for t < 0). It follows that i) H q (P, F (r)) = 0 for q ≥ 1, r ≥ 0; ii) all differentials dt = 0 for t ≥ 2, except that when k = n, r = 0 we have a short exact sequence n 1 F0 −→ H 0 (P, F ) → 0; 0 → Fn∗ −→

d

d

iii) we have an exact sequence induced by the maps d1 0 → Fk ⊗ S −k+r V → Fk−1 ⊗ S −k+1+r V → . . . → F1 ⊗ S −1+r V → F0r ⊗ S r V → H 0 (P, F (r)) → 0 where F0r = F0 for r ≥ 1 and F00 = F0 /dn Fn∗ in case k = n, r = 0. The lemma is now clear. To prove the proposition we define the graded S-module F = q Fq where Fq = H 0 (P, F (q))

q = 0 and

F0 = H 0 (P, F ) ⊕ Fn . Then from the long exact sequence in iii) just above we infer that F has a resolution ψn

ψn−1

ψ1

ψ0

0 → Fn −−→ Fn−1 −−−→ . . . F1 −→ F0 −→ F → 0 where deg ψ0 = 0 and deg ψi = 1 for i ≥ 1. In this way every involutive sheaf is the characteristic sheaf of an involutive module. As discussed at the end of §4 above, any involutive module M has a unique decomposition M = N ⊕ H0,n(M ) where Mq = Nq

q ≥ 1 and

H0,n (N ) = 0. The composite map involutive involutive involutive → → modules sheaves modules is given by M → N.

310


This is a consequence of statement iii) in the lemma above. The above proof and discussion have the following corollary: (110)

i) If F is an involutive sheaf, then the sheaf Euler characteristic satisfies χ(P, F (q)) = H 0 (P, F (q)),

(111)

q ≥ 0.

ii) If M is an involutive module with characteristic sheaf M, then M = q≥0 H 0 (P, M(q)) ⊕ H0,n(M )

(112)

iii) If M is an involutive sheaf, then the maps H 0 (P, M) ⊗ S q V → H 0 (P, M(q)) are surjective for q ≥ 0. We are now ready to prove a number of results that have been previously stated above. We begin with Proof of (96) above. Let M be an involutive sheaf and set W ∗ = H 0 (P, M) V = H 0 (P, O(1)) A∗ = H 0 (P, M(1)). Define

µ : W ∗ ⊗ V → A∗

to be the mapping on global sections induced by the sheaf mapping M ⊗O O(1) → M(1). By iii) above µ is surjective, and setting B = ker µ we have an exact vector space sequence µ → A∗ → 0. 0 → B → W∗ ⊗ V − The dual of this is

0→ A → W ⊗V∗ − → B ∗ → 0, σ

where A is our desired tableau and σ is a symbol mapping. By our discussion, if we start with the symbol σ1 of an involutive tableau, construct the associated characteristic sheaf, and then construct from this characteristic sheaf the symbol σ as above, σ1 differs from σ by a trivial symbol; i.e., σ1 is equivalent to σ in the sense of Definition 4.5 above. Proof of Theorem 3.6 in Chapter V. Let A ⊂ W ⊗V ∗ be an involutive tableau with characters s1 , . . . , sn and 1st prolongation A(1) ⊂ W ⊗ S 2 V ∗ . Then, respectively by definition and by Cartan’s test, (1130 )

dim A = s1 + s2 + · · · + sn

(1131 )

dim A(1) = s1 + 2s2 + · · · + nsn .

§5. Localization; the Proofs of Theorem 3.2 and Proposition 3.8 (1)

311

(1)

The characters s1 , . . . , sn of A(1) are uniquely determined by the relations (1)

(1)

(1)

s1 + · · · + sk = dim A(1) − dim Ak ,

(1141 )

k = 0, . . . , n.

It follows that (1)

dim A(1) = s1 + · · · + s(1) n ,

(1151 )

and also, as noted in equation (9) above, s(1) n = sn (1)

sn−1 = sn + sn−1 (1161 )

.. . (1)

s1 = sn + sn−1 + · · · + s1 . Note that if A has character l and Cartan integer σ = sl , then A(1) has the same character and Cartan integer. This will remain true for all the prolongations. Now, by Proposition 2.5 above, A(1) ⊂ W ⊗ S 2 V ∗ is again an involutive tableau (of order one) with 1st prolongation A(2) ⊂ W ⊗ S 3 V ∗ . By Cartan’s test (1)

(1)

dim A(2) = s1 + 2s2 + · · · + ns(1) n , which by (1161 ) gives (1132 )

dim A(2) = s1 + 3s2 + · · · + (n(n − 1/2)sn−1 + (n(n + 1)/2)sn . (2)

(2)

The characters s1 , . . . , sn are given by formula (1142 ), and then the relations (1152 ) and (1162 ) are valid. In general we will use (113q )–(116q ) to denote the (q) formulas (113)–(116) corresponding to A(q) . In general the characters si of (q) q+1 ∗ ⊂ W ⊗S V are given by (114q ) and then (115q ) and (116q ) are satisA fied. Recursively we therefore obtain our main formula (113q )

(q)

dim A

=

n k+q−1 q

k=1

sk .

Now let MA be the involutive module associated to A with characteristic sheaf MA . Then for q ≥ 1, by definition and by (112) A(q)∗ ∼ = (MA )q+1 = H 0 (P, M(q + 1)). Combining this with (111) and (113q ) gives the following expression for the sheaf Euler characteristic n k+q−1 sk . χ(P, M(q + 1)) = q k=1

312


To complete the proof we shall compare the two sides of this formula for large q. If the tableau A has character l and Cartan integer σ, thus sl = σ and sl+1 = · · · = sn = 0, then (117)

n k+q−1 q

k=1

sk =

σq l−1 + (lower order terms in q). (l − 1)!

On the other hand we may evaluate the sheaf Euler characteristic χ(P, M(q + 1)) by the Riemann–Roch formula (Fulton and Lang [1985]). (Of course, there are more elementary methods but this is perhaps the clearest conceptually.) If supp M = Ξ then it follows from that formula that χ(P, M(q + 1)) = P (M, q) Ξ

where P (M, q) involves the Chern classes of M and ω = c1 (O(1)) as they appear in the Chern character of M(q) twisted by the Todd class of Ξ. Explicitly (118)

P (M, q) =

κq m−1 ωm−1 + (lower order terms in q) (m − 1)!

where m − 1 = dim Ξ and κ is the fibre dimension of M over a general point of Ξ. From the preceeding four equations we infer first that m = l and secondly that σ = κ ωm−1 = κδ Ξ

where δ = deg Ξ.

We now turn to the proof of Proposition 3.10. From Corollary 3.8 what must be proved is this:3 (119)

Let M be a quotient of a free module W ∗ ⊗ SV . Then there exists a p0 , depending on dim W ∗ and the Hilbert function PM (q) of M , such that the Koszul homology groups Hp,q (M ) = 0 for p ≥ p0 , 0 ≤ q ≤ n = dim V.

Referring to the discussion following the proof of Proposition 4.3 (cf. (i) and (iii) in that discussion), the statement (119) is an assertion about the minimal free resolution of M . More precisely, it is equivalent to (120)

Bp,q = 0

if q − p ≥ p0 .

3 We would like to thank Mark Green for help with this argument. His paper Green [1989a] serves as a general reference on Koszul groups, and Green [1989b] has a discussion related to finding an effectively computable bound for p0 . A weaker version of this appears in Goldschmidt [1968a] and in Goldschmidt [1974].


313

If R is the module of relations defined by 0 → R → W ∗ ⊗ SV → M → 0, E0 then clearly (120) for M implies the same statement for R, so that (119) follows from (and is in fact equivalent to) the assertion: (121)

Let R ⊂ W ∗ ⊗ SV be a sub-module of a free module. Then there is a p0 , depending only on dim W ∗ and the Hilbert function PR(q) = dim W ∗ . PSV (q) − PM (q), such that Hp,q (R) = 0 for

p ≥ p0 , 0 ≤ q ≤ n.

We will prove (121) by a localization argument. Given a graded module we ¯ = ⊕R ¯q, R ¯ q = H 0 (R(q)) the module denote by R the corresponding sheaf and by R ¯ is usually called the saturation of R. There is always a associated to the sheaf R. R ¯ and when R is a submodule of a free module this mapping module mapping R → R, is injective. Proof. Suppose that R ⊂ W ∗ ⊗ SV . Then R ⊂ W ∗ where W ∗ is the trivial bundle ¯ ⊂ W ∗ ⊗ SV , while on the other with fibre W ∗ . It follows on the one hand that R hand we infer from (109) that W ∗ ⊗ SV = W ∗ ⊗ SV . From the commutative diagram 0 ↓ 0 → R → W ∗ ⊗ SV ↓ ↓≈ ∗ ¯ 0 → R → W ⊗ SV ¯ is injective. we infer that R → R

The main step in our proof of (121) is to show that a weaker statement is true ¯ is saturated. More precisely, recall that by definition the Hilbert when R = R polynomial χR (q) of a graded module R is the unique polynomial such that χR (q) = PR (q) for large q. Then it is clear that χR (q) =

(−1)i dim H i (R(q))

i

is the Euler characteristic of the twists of the localization of R. The weaker version of (121) is this: (122)

Let R ⊂ W ∗ ⊗ SV be a saturated sub-module of a free module. Then there exists a p0 , depending only on dim W ∗ and on the coefficients of the Hilbert polynomial χR (q), such that

314


Hp,q (R) = 0 for

p ≥ p0 , 0 ≤ q ≤ n.

Assuming this result we will complete the proof of (121), and then we shall prove (122). We consider the exact sequence ¯ → R/R ¯ 0→R→R → 0.

(123)

¯ Since R/R is a finite dimensional vector space over C, it is clear that χR (q) = χR¯ (q). ¯ to conclude that Thus we may apply (122) to the saturated module R (124)

¯ = 0 for p ≥ p1 = p1 (dim W ∗ , PR (q)), 0 ≤ q ≤ n Hp,q (R)

where p1 depends on dim W ∗ and χR¯ (q) = χR (q). On the other hand, let p2 be an integer such that PR (q) = PR¯ (q) for q ≥ p2 . We will see below that there is a q0 depending only on χR¯ (q) = χR (q), and therefore only on PR (q), such that χR¯ (q) = PR¯ (q) for q ≥ q0 . Thus p2 may be chosen to satisfy p2 ≥ q0 and PR (q) = χR (q) for q ≥ p2 , and therefore p2 = p2 (PR (q)) may be assumed to also depend only on the Hilbert function of R. Then ¯ (R/R) p = 0 for

p ≥ p2 (PR (q)).

But then clearly ¯ Hp,q (R/R) = 0 for p ≥ p2 (PR (q)).

(125)

Combining (124) and (125) and using the long exact homology sequence of (123) gives the desired result (121). It remains to give the Proof of (122). We shall use the following. Definition 5.7. A coherent sheaf R on P = PV ∗ is said to be m-regular in case H i (R(m − i)) = 0 for i > 0. The smallest m such that R is m-regular is called the regularity m(R) of R. Lemma 5.8. Let

0 → En → En−1 → · · · → E0 → R → 0 Ep = q Bp,q ⊗ S [−q]

be the minimal resolution of a saturated module R. Then for the localization R the regularity m(R) = max{q − p : Bp,q = 0}. Proof. We will prove the result in the first non-trivial case when the minimal resolution has two terms; the general case is proved by an analogous argument with


315

spectral sequences replacing the long exact cohomology sequence—cf. Theorem 2.3 in Green [1989a]. We first note that the sheaf O on Pn−1 is 0-regular but is not (−1)-regular; this follows from (109) above. In general, by using shifts we may reduce to considering the case of 0-regularity. Suppose that (125)

0→

p

ϕ

B1,p ⊗ S[−p] − →

q

B0,q ⊗ S[−q] → R → 0

is the minimal resolution of R. Writing ϕ = p,q ϕp,q where ϕp,q : B1,p ⊗ S[−p] → B0,q ⊗ S[−q] is given by a matrix of homogeneous polynomials of degree p − q, it follows trivially that ϕp,q = 0 for p < q and by minimality that ϕp,p = 0. Suppose first that max{q − p : Bp,q = 0} = 0. Then by localizing (125) we have (126)

0→

p

B1,p ⊗ O(−p − i) →

q

B0,q ⊗ O(−q − i) → R(−i) → 0

where B0,q = 0 for q > 0, B1,p = 0 for p > 1. Then only non-trivial piece of the exact cohomology sequence is (127)

n−1 (B1,p ⊗ O(−p − i)) → 0 → H n−2 (R(−i)) → p H n−1 (B0,q ⊗ O(−q − i)) → H n−1 (R(i)) → 0 q H

Taking i = n − 2 we have H n−1 (B1,p ⊗ O(−p − n − 2)) = 0 since B1,p = 0 for p > 1, and taking i = n − 1 we have H n−1 (B0,q ⊗ O(−q − (n − 1)) = 0 since B0,q = 0 for q > 0. Thus R is 0-regular. Conversely, suppose that R is 0-regular. From the lemma below we see that R is m-regular for m ≥ 0. Let p0 be the largest integer such that B1,p0 = 0. Taking i = n − p0 in (127) and using that ϕp0 ,q = 0 for q ≥ p0 we obtain ∼

0 → H n−2 (R(−n + p0 )) −→ B1,p0 ⊗ H n−1 (O(−n)) → 0. For p0 ≥ 2 we have −n+p0 = −(n−2)+m where m ≥ 0 and so H n−2 (R(−n+p0 )) = 0. Thus B1,p = 0 for p ≥ 2. Similarly, let q0 be the largest integer such that B0,q = 0. If q0 ≥ 1 then take i = n − q0 in (127) and use that ϕp,q0 = 0 for p ≥ 1 to obtain ∼

0 → B0,q0 ⊗ H n−1 (O(−n)) −→ H n−1 (R(−n + q0 )) → 0. Writing −n + q0 = −(n − 1) + m where m ≥ 0, this last term is zero by the mregularity of R. Thus m(R) = 0 as desired. Referring to (iii) in the discussion following the proof of Proposition 4.2, we have (128)

m(R) = max{p : Hp,q (R) = 0 for some q}.

Thus our desired result (122) follows from the following.

316


Proposition 5.9. Let R ⊂ W ∗ be a coherent subsheaf of the trivial sheaf W ∗ ∗ on P . Then there exists a polynomial f(s, ai ) that depends on s = dim W and q such that the ai the coefficients ai of the Hilbert polynomial χ(R(q)) = i regularity m(R) ≤ f(s, ai ). Proof (cf. Mumford [1966], Chapter XIV). We shall give the argument when R ⊂ O; the general case is the same. Then R is a sheaf of ideals on P ∼ = Pr−1 . The proof will be by induction, and we choose a general hyperplane H ⊂ P defined by h ∈ H 0 (OP (1)). Then we have the sequence h

→ R → R ⊗OP OH → 0 0 → R(−1) − RH which is injective on the left. A local argument shows that the sequence is in fact exact and that RH ⊂ OH is a subsheaf. Tensoring with OP (m + 1) we obtain (129)

0 → R(m) → R(m + 1) → RH (m + 1) → 0,

and therefore χ(RH (m + 1)) = χ(R(m + 1)) − χ(R(m)) n−1 # m + 1 m $ − = ai i i i=0

=

n−2 i=0

ai+1

m i

.

The induction assumption applies to RH , so we may assume that it is g(ai )-regular for a suitable polynomial g depending only on n; we set m1 = g(ai ). Then the exact cohomology sequences of (129) give ρm+1

(130 )

0 → H 0 (R(m)) → H 0 (R(m + 1)) −−−→ H 0 (RH (m + 1)) → H 1 (R(m)) → H 1 (R(m + 1)) → 0

for m ≥ m1 − 2, and (130 )

0 → H i (R(m)) → H i (R(m + 1)) → 0

for i ≥ 2, m ≥ m1 − 2. Since H i (R(m + 1)) = 0 for m sufficiently large and i ≥ 1 this last sequence gives (131)

H i (R(m)) = 0

i ≥ 2, m ≥ m1 − 2.

Turning to H 1 , from (130 ) we see that if m ≥ m1 − 2 then either ρm+1 is surjective or else h1 (R(m)) > h1 (R(m + 1)). Since h1 (R(m)) < ∞ there is an m2

§6. Proof of Theorem 3.15 in Chapter V; Guillemin’s Normal Form

317

such that m2 ≥ m1 and ρm2 is surjective. In a moment we will show that ρm2 +1 is also surjective, from which it follows that, for m ≥ m1 , h1 (R(m)) is strictly decreasing as a function of m until it reaches zero. Then clearly R is [m1 + h1 (R(m1 ))]-regular. But h1 (R(m1 )) = h0 (R(m1 )) − χ(R(m1 )) ≤ h0 (OP (m1 )) − χ(R(m1 )) = h(ai , m1 ) = f(ai ) where h and f are suitable polynomials. It remains to show that ρm2 +1 is surjective. By the lemma below, H 0 (OP (1)) ⊗ H 0 (RH (m2 )) → H 0 (RH (m2 + 1)) is surjective. Thus the horizontal composite map in the commutative diagram H 0 (OP (1))⊗H 0 (R(m2 ))

1⊗ρm

2 −−−−→

H 0 (OP (1))⊗H 0 (RH (m2 ))→H 0 (RH (m2 +1))

↓ ρm2 +1

H 0 (R(m2 +1))

is surjective, and it then follows that ρm2 +1 is surjective. We will be done once we prove the Lemma 5.10. If R is m-regular, then for k > m (i) H 0 (OP (1)) ⊗ H 0 (R(k)) → H 0 (R(k + 1)) is onto; and (ii) R is k-regular. Proof. From the exact sequence (129) for m − i − 1 we obtain H i (R(m − i)) → H i (RH (m − i)) → H i+1 (R(m − i − 1)),

and so RH is m-regular. We may apply an induction hypothesis to conclude (i) and (ii) for RH . From the exact sequence (129) for m − i H i+1 (R(m − i − 1)) → H i+1 (R(m − i)) → H i+1 (RH (m − i)),

i ≥ 0,

and from (ii) for RH we see that R is (m + 1)-regular. From the diagram H 0 (R(k−1))⊗H 0(OP (1))

α

− →

H 0 (RH (k−1))⊗H 0 (ϕP (1))→H 1 (R(k−2))⊗H 0(OP (1)) ↓β

µ 0

H (R(k−1))

− → h

0

H (R(k))

− →

0

H (RH (k))

we see that α is surjective for k > m, and by (i) for RH , β is also surjective for k > m. But then H 0 (R(k)) is spanned by image µ + image h, while clearly image h ⊆ image µ.

318


§6. Proof of Theorem 3.15 in Chapter V; Guillemin’s Normal Form. It is clear that Theorem 3.15 in Chapter V is a purely algebraic result dealing with a tableau A ⊂ W ⊗ V ∗ . Accordingly, we shall reformulate the result in a purely algebraic manner and then prove the reformulated version. Setting U = W ⊗ V ∗ /A we consider the symbol mapping σ : W ⊗V∗ → U and define (132)

Ξp = {Ω ∈ Gn−p (V ∗ ) : σ : W ⊗ Ω → U fails to be injective} = {Ω ∈ Gn−p (V ∗ ) : A ∩ W ⊗ Ω = 0}.

It is clear that

Ξn−1 = ΞA ⊂ PV ∗

is the characteristic variety of the tableau. Using the projective duality isomorphism ≈

Gn−p(V ∗ )−→Gp (V ) ↓ ↓ Ω we define in the obvious way

→ Ω⊥

Ξ⊥ p ⊂ Gp (V ).

On the other hand, in Definition 3.14 in Chapter V we have defined Λp ⊂ Gp (V ), and we shall show that: (133)

If the tableau A has character l, then Ξ⊥ p = Λp for l ≤ p ≤ n − 1.

Proof. We let {wa }, {vi } and zε = Aaεi wa ⊗ vi∗ be bases for W , V , and A respectively. Setting πia = Aaεi zε∗ ∈ A∗ we consider the tableau matrix

π=

π11 . . . πn1 .. . . π1s . . . πns


319

Using the additional index ranges 1 ≤ λ, µ ≤ l l + 1 ≤ ρ, σ ≤ n, the character l of the tableau A is the smallest integer with the following property: If the basis {vi } is chosen generically, then (134)

πρa ≡ 0 modulo{πλb }.

That is, the πρa ∈ A∗ in the last n − l columns of π should be linear combinations of the πλb in the first l columns. Clearly (134) is equivalent to (135)

all πλb (ψ) = 0 ⇒ all πρa (ψ) = 0, for any ψ ∈ A.

Now let Ω ∈ Gn−l (V ∗ ) and choose a basis {vi } for V so that Ω = span{vρ∗ }, or equivalently

Ω⊥ = span{vλ }.

Recalling that if s1 , . . . , sn denote the characters of A and s(Ω) denotes the dimension of span{πλb } ⊂ A∗ , Ω ∈ Λl ⇔ s(Ω) < s1 + · · · + sl ⇔ (134) fails to be true there exists 0 = ψ ∈ A with ⇔ . all πλb (ψ) = 0 but some πρa (ψ) = 0 But then (136)

0 = πia (ψ)wa ⊗ vi∗ ∈ A ∩ W ⊗ Ω.

Conversely, any element of A is of this form for a suitable ψ and the condition that (136) lie in W ⊗ Ω is that all πλb (ψ) = 0. This proves (133) when p = l, and it is clear that the same argument works for l ≤ p ≤ n − 1. We note that (133) does not depend on the involutivity of A. Using (133), it is a small and straightforward exercise to reformulate Theorem 3.15 in Chapter V as follows: Theorem 6.1. For an involutive tableau A ⊂ W ⊗ V ∗ of character l, the following are equivalent conditions on Ω ∈ Gn−l (V ∗ ): i) A ∩ W ⊗ Ω = 0 ii) for some line Lξ ⊂ Ω A ∩ W ⊗ Lξ = 0. Here, Lξ is the line in V ∗ corresponding to [ξ] ∈ PV ∗ , and A ∩ W ⊗ Lξ = 0 is just our condition (132) that [ξ] ∈ ΞA . On the other hand, A ∩ W ⊗ Ω = 0 is the condition (132) that Ω ∈ Ξl . It is clear that even without assuming that A is

320


involutive, ii) ⇒ i), and what we have to do is to show that i) ⇒ ii). This requires involutivity and amounts to showing that If A ∩ W ⊗ Ω = 0, then A ∩ W ⊗ Ω contains a non-zero decomposable vector w ⊗ ξ. Equivalently, we must show that: PΩ ∩ ΞA = ∅ ⇒ σ : W ⊗ Ω → U is injective.

(137)

We shall formulate a stronger result than (137), and shall then prove this stronger result by a localization argument. Let M be an involutive graded S-module with canonical resolution 0 → En → En−1 → · · · → E1 → E0 → M → 0 having localization ψn−1

ψn

0 → En (−n) −−→ En−1 (−(n − 1)) −−−→ · · · →

(138)

ψ1

ψ0

E1 (−1) −→ E0 −→ M → 0.

Here M is the characteristic sheaf of M and Ei is the trivial vector bundle with fibre Ei = H0,i(M ) where Ei = Ei ⊗ SV . For any k-plane Ω ∈ Gk (V ∗ ) we may restrict all sheaves to PΩ ∼ = Pk−1 , and then H 0 (PΩ, Ei (q)) ∼ = E i ⊕ S q Ω∗ . Hence there are induced maps ψi

Ei ⊗ S q Ω∗ −→ Ei−1 ⊗ S q+1 Ω∗ ,

q ≥ 0.

Proposition 6.2. If PΩ ∩ supp M = ∅, then for each l ≥ 1 the sequence El → El−1 ⊗ Ω∗ → · · · → E1 ⊗ S l−1 Ω∗ → E0 ⊗ S l Ω∗ → 0

(139) is exact.4

The case l = 1 is the exact sequence E1 → E0 ⊗ Ω∗ → 0. In case M = MA is the graded module associated to be involutive tableau A, then supp M = ΞA is the characteristic variety of the tableau and the above sequence is σ∗

B −→ W ∗ ⊗ Ω∗ → 0, which dualizes to

σ

0→W ⊗Ω− → U. Thus, Proposition 6.2 in the case l = 1 implies (137). 4 The

l used in this proof has nothing to do with the character of the tableau A.


321

Proof of Proposition 6.2. We set Ξ = supp M and will prove the exactness of (139) first where PΩ is a point in V ∗ and then in general. dim Ω = 1. In this case PΩ is a point [ξ] ∈ PV ∗ \Ξ. Denoting by Oξ the local ring of OPV ∗ at [ξ], the exact stalk sequence of (138) at [ξ] is the following exact sequence of Oξ -modules 0 → En (−n)ξ → En−1 (−(n − 1))ξ → · · · → E1 (−1)ξ → E0ξ → 0.

(140)

Here we are using that Mξ = 0 if [ξ] ∈ / supp M. Now Ei (−q) is a vector bundle whose fibre over [ξ] ∈ PV ∗ is Ei ⊗ L−q ξ . Thus Ei (q)ξ = Ei ⊗C L−q ξ ⊗C Oξ , and from Nakayama’s lemma it follows that (140) gives the exact sequence of vector spaces −(n−1)

0 → En ⊗ L−n → En−1 ⊗ Lξ ξ

→ · · · → E1 ⊗ L−1 ξ → E0 → 0.

Tensoring this sequence with Llξ gives the exactness of (139) for all l when dim Ω = 1. This case is due to Quillen [1964]. According to the discussion centered around (87) and (88) above it may be rephrased as follows: for non-characteristic covectors, the symbol sequence of the Spencer sequence associated to an involutive linear P.D.E. system is exact. Remark. If we set hi = dim Ei = dim H0,i (M ), then from

−1 E2 ⊗ L−2 ξ → E1 ⊗ Lξ → E0 → 0

we infer that (141)

h2 ≥ h1 − h0 .

This inequality, a special case of which is due to Cartan (cf. the reference in footnote 2), has the following interpretation for an involutive, constant coefficient linear P.D.E. system (142)

Baλi

∂ua (x) = f λ (x). ∂xi

Here 1 ≤ a ≤ s = number of unknown functions and 1 ≤ λ ≤ t = number of equations. We have h0 = s, h1 = t,

322


and so if the system (142) is determined or underdetermined then (141) doesn’t say anything. However, suppose we are in the overdetermined case t > s. Setting

Dλ = Baλi wa∗ ⊗ ∂/∂xi ,

h2 is the number of linearly independent, 1st order constant coefficient relations (cf. (86) above) mjλ ∂/∂xj Dλ = 0.

(143) From (141) we have that:

The number of relations (143) is ≥ t − s. As discussed in §4 above, the relations (143) give the 1st order compatibility conditions ∂f λ (x) mjλ =0 ∂xj for the formal solvability of (142). By involutiveness, all the compatibility conditions are 1st order. Returning to the proof of Proposition 6.2 in general, we suppose that dim Ω = k so that PΩ ∼ = Pk−1 . Over PΩ we have the exact complex (144)

0 → En (−n + l) → En−1 (−n + l + 1) → · · · → E1 (l − 1) → E0 (l) → 0

where Ek (m) is Ek ⊗ OPΩ (m). For −n ≤ j ≤ 0 we set Fj = E−j (j + l) so that (144) becomes the exact complex5 (145)

0 → F−n → F−n+1 → · · · → F−1 → F0 → 0.

We note that by (109) above

(146)

⎧ ⎪ ⎨0 q H (PΩ, Fj ) = 0 ⎪ ⎩ 0

q = 0, k − 1 q = k − 1, j ≥ −k − l − 1 . q = 0, j ≤ −l − 1

Associated to the complex of sheaves (145) are two spectral sequences both abutting to the hypercohomology of the complex of sheaves. Since (145) is exact one of these has E1 term equal to zero. Thus the other spectral sequence abuts to zero. Now by our indexing convention both spectral sequences are in the second quadrant. The one with non-zero E1 term has E1p,q = H q (PΩ, Fp ) 5 This argument is similar to the proof of Proposition 5.5 above except that we have chosen to use a different indexing convention.


323

where −n ≤ p ≤ 0 and 0 ≤ q ≤ k − 1. By (146) the only non-zero terms are

E1p,k−1 , p ≤ −k − l E1p,0 ,

p ≥ −l.

From this we see that the only possible non-zero differentials are d1 : E1p,0 → E1p+1,0 d1 : dk :

p ≥ −l

E1p,k−1 → E1p+1,k−1 Ek−k−l,k−1 → Ek−l,0 .

p ≤ −k − l − 1

It follows that E2p,q = 0 unless

p = −k − l, p = −l,

q=k−1 q=0

E2 = · · · = Ek Ek+1 = 0. In particular, the complex d

1 H 0 (PΩ, F−l ) −→ H 0 (PΩ, F−l+1 ) → · · · → H 0 (PΩ, F−1 ) → H 0 (PΩ, F0 ) → 0

is exact. This implies Proposition 6.2.

324


Guillemin’s Normal Form. Victor Guillemin established a “normal form” for the Spencer complex of an involutive linear P.D.E. system (cf. Guillemin [1968]). This result is closely related to the special case l = 2 of Proposition 6.2 and so we should like to discuss it here. Let A ⊂ W ⊗V ∗ be an involutive tableau, considered as the image of an injective linear mapping π : A → W ⊗ V ∗. In terms of bases {wa} for W and {xi } for V ∗ we write π = πia wa ⊗ xi where πia ∈ A∗ . We will establish a certain normal form for the symbol relations on the πia ’s. This normal form will in fact correspond to the symbol relations when the tableau is put in the normal form given by equation (90) in Chapter IV. Referring to that discussion, the forms πia for a < si are given by linear equations (147)

πia =

aj b Bib πj .

b≤sj j≤i

Involutivity of the tableau will have strong commutation properties on matrices aj above. For example, when s1 = · · · = sl = s and sl+1 = derived from the Bib · · · = sn = 0, the above equations reduce to (147 )

aλ b πρa = Bρb πλ

where 1 ≤ λ ≤ l and l + 1 ≤ ρ, σ ≤ n. We have seen in section 5 of Chapter IV that involutivity is equivalent to the commutation relations [Bρ (ξ), Bσ (ξ)] = 0 for all ξ where aλ Bρ (ξ) = Bρb ξλ .

There will be an analogous statement in the general case, one that will be given following a general discussion. The geometric picture is this: If A has character l then the complex characteristic variety ΞA has dimension l − 1, and therefore by a generic linear projection may be realized as a finite branched covering over a Pl−1 . We shall then explore how this representation may be used to at least partially normalize the relations (147). To carry this out we let Ω ⊂ V ∗ be a maximal, non-characteristic subspace. Then dim V ∗ = n, dim Ω = n − l and PΩ ∩ ΞA = ∅. Set E = Ω⊥ ⊂ V so that E ∗ ∼ = V ∗ /Ω. Then PE ∗ ∼ = Pl−1 and by linear projection there is a diagram PV ∗ \PΩ ⊃ ΞA ↓ω ˜ ↓ω ˜ PE ∗ = PE ∗ .


325

ω ˜

where the map ΞA − → PE ∗ is a finite branched covering map. If ξ ∈ PE ∗ and we set (the following has intrinsic meaning) {Ω, ξ} = span Ω, ξ, then {Ω, ξ} ∼ = Pn−l and we have

ω ˜ −1 (ξ) = {Ω, ξ} ω ˜ −1 (ξ) = {Ω, ξ} · ΞA .

Choose our basis {xi } for V ∗ so that Ω = span{xl+1 , . . . , xn }. We keep the additional index ranges 1 ≤ λ, µ ≤ l l + 1 ≤ ρ, σ ≤ n and write points in V ∗ as ζ = (ξ, η) = ξλ xλ + ηρ xρ ; i.e., in homogeneous coordinates ζ = [ξ1 , . . . , ξl ; ηl+1 , . . . , ηn ] (from now on we drop the bracket around points ξ in a projective space). Then with this notation ω ˜ (ζ) = ξ. Since Ω is non-characteristic, we have (148)

A ∩ W ⊗ Ω = 0.

Elements in W ⊗ Ω are of the form ψ = ψρa wa ⊗ xρ . Thus (148) is equivalent to (cf. (135) above) πλa (ψ) = 0 ⇒ πρa (ψ) = 0, ψ ∈ A. In other words, the basis {xi } for V ∗ has the property that Ω = span{xl+1 , . . . , xn } is non-characteristic if, and only if, the forms πia are all linear combinations of the πλa for 1 ≤ λ ≤ l, in which case we have the relations (147) among the symbol relations. We now recall from §5 of Chapter IV that, in the involutive case, we may assume that πia = 0 for a > s1 . Thus we may as well assume that s1 = s, and then the equations (147) give in particular relations (147 ). We shall derive a general commutation property of the matrices Bρ (ξ). For this we let ξ ∈ E ∗ and consider expressions (149)

w ⊗ ξ + wρ ⊗ xρ ∈ A ∩ (W ⊗ {Ω, ξ}).

326


We note that, by (148), the tensor wρ ⊗ xρ ∈ W ⊗ Ω is uniquely determined by the property that w ⊗ ξ + wρ ⊗ xρ ∈ A, and we define Wξ = {w ∈ W : there exists an expression (149)}. Clearly Wξ ⊂ W is a linear subspace, and we shall show that for any tensor (149) Bρ (ξ)w = wρ

(150)

where Bρ (ξ) is the linear transformation associated to the matrix (147 ). Proof. If w = µa wa ∈ Wξ , then for some ψ ∈ A we have π(ψ) = w ⊗ ξ + wρ ⊗ xρ = µa ξλ wa ⊗ xλ + µaρ wa ⊗ xρ where we have set wρ = µaρ wa and the ξλ are the components of ξ. This gives πλa (ψ) = µa ξλ πρa (ψ) = µaρ . But then (147 ) gives aλ b µ ξλ , µaρ = Bρb

i.e., Bρ (ξ)w = wρ . We shall now show that Proposition 6.3. With the above notations (151)

Bρ (ξ)Wξ ⊆ Wξ

(152)

[Bρ (ξ), Bσ (ξ)]|Wξ = 0.

Proof. We shall use the exact sequence (139), or rather its dual. We recall the dual sequence (88) in §3 above of the canonical resolution of an involutive module, which using the definition of the Spencer cohomology groups may be written as δ¯

0 → A(q) → W ⊗ S q+1 V ∗ − →

W ⊗V∗ A

δ¯

→ ⊗ Sq V ∗ −

W ⊗ Λ2 V ∗ δ(A ⊗ V ∗ )

⊗ S q−1 V ∗ .

Here, for ψ ∈ W ⊗ Λi V ∗ and P ∈ S q+1−i V ∗ the mapping δ¯ is given by (153)

¯ ⊗ P) = δ(ψ

i

ψ ∧ dxi ⊗

∂P ∂xi

mod δ(A ⊗ Λi V ∗ )


327

(cf. the proof of (90) in §3 above). For Ω ⊂ V ∗ a non-characteristic subspace, the dual of (139) above when l = 2 is # $ W ⊗V∗ W ⊗ Λ2 V ∗ δ¯ δ¯ 2 (154) 0 → W ⊗S Ω− → → ⊗Ω − . A δ(A ⊗ V ∗ ) By (153) we have for ψ ∈ W ⊗ Λi V ∗ and P (x) depending only on the last (n − l)variables, i.e., P ∈ SΩ, ¯ ⊗ P (xl+1 , . . . , xn )) = δ(ψ

ψ ∧ dxρ ⊗

ρ

∂P . ∂xρ

Suppose now that w ⊗ ξ + wρ ⊗ xρ ∈ A ⊂ W ⊗ V ∗ . Taking the exterior product with ξ gives wρ ⊗ xρ ∧ ξ ≡ 0 mod δ(A ⊗ V ∗ ) (here we are using (153) above when P = ξ). But ¯ ρ ⊗ ξ) ⊗ xρ ) wρ ⊗ xρ ∧ ξ = −δ((w where (wρ ⊗ ξ) ⊗ xρ ∈ W ⊗ V ∗ ⊗ Ω. By the exactness of (154) we have # $ 1 W ⊗V∗ ρ ρ σ ¯ (wρ ⊗ ξ) ⊗ x = δ( wρσ x x ) ∈ ⊗Ω 2 A where wρσ = wσρ ∈ W. It follows that

wρ ⊗ ξ ≡ wρσ xσ modulo A . wρσ = wσρ

The first of these equations gives (151) and the second gives (152).

We may now complete our description of the branched covering ω ˜ : ΞA → PE ∗ . ∗ Given ξ ∈ P E there are points ˜ −1 (ξ). [ξ, η(ξ)] ∈ ΞA ∩ ω Here, we are writing η(ξ) to express the fact that via the finite branched covering mapping ΞA → PE ∗ , the inverse image of a point ξ ∈ PE ∗ is a finite number of points [ξ, η(ξ)] whose η-coordinates are algebraic functions of ξ. Thus, setting ζ = (ξ, η(ξ)) there is a non-zero vector w ∈ W with σζ (w) = 0. This gives that w ⊗ ξ + w ⊗ η(ξ) ∈ A ∩ {Ω, ξ},

328


and so w ∈ Wξ . Referring to (150), we have Bρ (w) = ηρ (ξ)w where the ηρ(ξ) are the components of η(ξ). Thus, the commuting linear transformations Bρ (ξ)|Wξ may be put in Jordan normal form with the last n−l coordinates on each sheet giving the eigenvalues for the common eigenvector of these transformations. At this stage, one needs to be careful in treating the non-semisimple parts of the Bρ (ξ)’s. Moreover, one must worry about the characteristic ideal and not just the characteristic variety.6 We refer to Gabber [1981] for further references and discussion. §7. The Graded Module Associated to a Higher Order Tableau. It is well known that a system of higher order P.D.E. may be rewritten as a (much larger) 1st order P.D.E. system. In general, however, it is preferable to treat the higher order system directly. Similarly, in many geometric examples the tableau of a linear Pfaffian differential system looks like the tableau of a higher order system—we have called these a tableau of order p. It is desirable to adapt the formalism—characteristic variety, graded module associated to a tableau, etc.— to a tableau of order p, and this is what we shall do in this section. For reasons to appear below, we shall make the notation shift p → q − 1. We consider a tableau of order q − 1 A ⊂ W ⊗ Sq V ∗ with prolongations

A(k) ⊂ W ⊗ S k+q V ∗

where (cf. (2)–(4) above) (155)

A(k) = W ⊗ S k+q V ∗ ∩ A ⊗ S k V ∗ .

We set A=

(k) k≥0 A

⊂ W ⊗ SV ∗ .

By definition the symbol of A is B = A⊥ ⊂ W ∗ ⊗ S q V, and the symbol module is, again by definition, B= where

k≥0 Bk

Bk = A(k)⊥ ⊂ W ∗ ⊗ S k+q V.

6 Again it is interesting to note in Cartan [1953] attention is drawn to the technical difficulties encountered when there is an essential non-diagonal piece to the Bρ (ξ)’s (cf. footnote 9 in Chapter V).

§7. The Graded Module Associated to a Higher Order Tableau

329

As before (the case q = 1), B is the SV -submodule of W ∗ ⊗ SV generated by B = B0 with an appropriate shift in grading. Example 7.1. We make the identification SV ∼ = C[∂/∂x1 , . . . , ∂/∂xn ] by the mapping vi → ∂/∂xi . Given a tableau A ⊂ W ⊗ S q V ∗ of order q − 1 with symbol B ⊂ W ∗ ⊗ S q V , we choose a basis Dλ = BaλI wa∗ ⊗ ∂ I /∂xI ,

|I| = q,

for B, and thereby establish a 1–1 correspondence between tableaux of order q − 1 and q th order, constant coefficient homogeneous linear P.D.E. systems Dλ u(x) = 0. The symbol module is just the sub-module of W ∗ ⊗ C[∂/∂x1 , . . . , ∂/∂xn ] generated by the Dλ ’s. Returning to the general discussion we recall our “shift” notations (W ∗ ⊗ SV )k

[q−1]

(W ∗ ⊗ SV )[q−1]

= W ∗ ⊗ S k+q−1 [q−1] = k≥0 (W ∗ ⊗ SV )k

and give the following: Definition 7.2. Given a tableau A ⊂ W ⊗ S q V ∗ , we define the associated graded module MA by the exact sequence (156)

0 → B → (W ∗ ⊗ SV )[q−1] → MA → 0,

where the 1st map is homogeneous of degree one and the 2nd is homogeneous of degree zero. It follows that MA = k≥0 (MA )k

where (MA )k =

A(k−1)∗ W ∗ ⊗ S q−1 V

k≥1 k = 0.

When q = 1 this coincides with the graded module associated to an ordinary tableau introduced in §3 above. Generalizing (57) above we have (157)

The tableau A ⊂ W ⊗ S q V ∗ is involutive if, and only if, the associated graded module MA is involutive in the sense of Definition 4.1 above.

The proof of (57) is based on (and, in fact, is equivalent to) the surjectivity of the (q+1) (q) maps ∂/∂xi : Ai−1 → Ai−1 , q ≥ 0. According to Cartan’s test, this surjectivity is by definition the same as involutiveness for a tableau of any order. Consequently,

330


the proof of (57) given above carries over pretty much verbatim to give a proof of (157). The “pretty much” refers to the fact that some care must be taken in the definition of the degree zero piece of MA . This will be discussed now. The simplest involutive module is (W ∗ ⊗ SV )[q−1] . However, this is not a free module if q ≥ 2. To explain this we remark that (W ∗ ⊗ SV )[q−1] corresponds to the empty P.D.E. system. However, when q ≥ 2 there are compatibility conditions that functions uα I , |I| = q − 1, be (q − 1)-jets; these conditions correspond to the non-freeness of (W ∗ ⊗ SV )[q−1] as expressed by7 H0,1 ((W ∗ ⊗ SV )[q−1] ) = 0 if q ≥ 2. The meaning of “correspond to” will be elaborated on below. Example 7.3. We consider a constant coefficient linear P.D.E. system for one unknown function (158)

B λI

∂ q u(x) = f r (x), ∂xI

λ = 1, . . . , t.

The symbol module B of (158) is generated by the homogeneous polynomials P λ (ξ) = B λI ξI ∈ S q V. If (158) is involutive, then it follows from the exact homology sequence of (156) that B is an involutive S-module. Let (159)

ϕ1

ϕ0

· · · → E1 −→ E0 −→ B → 0

be its canonical resolution where Ei = Ei ⊗C S and ϕ0 has degree zero while ϕi has degree one for i ≥ 1. The localization of (159) is (160)

ϕ1

ϕ0

· · · → E1 (−q − 1) −→ E0 (−q) −→ T → 0

where T ⊂ O is the sheaf of ideals of the characteristic variety Ξ = {ξ : P λ (ξ) = 0}. Over a point ξ ∈ / Ξ, the fibre sequence of (163) is −(q+1)

→ E1 ⊗ Lξ

→ E0 ⊗ L−q ξ → C → 0.

Setting hi = dim Ei (= dim H0,i(B)) it follows that h1 ≥ h0 − 1.

(161) Now

h0 = number of equations (158) while h1 is the number of independent linear compatibility conditions (162) 7 This

mjλ

q ∂ λI ∂ (B ) = 0. ∂xj ∂xI

discussion here is closely related to the truncation example discussed in §4 above.


331

Setting Qλ (ξ) = mjλ ξj (162) is equivalent to (163)

Qλ (ξ)P λ (ξ) = 0,

deg Qλ (ξ) = 1.

We therefore conclude from (164) that: (164)

If (158) is involutive, then there are at least t − 1 independent linear relations (163) among the polynomials P λ (ξ).

This result is due to Cartan in the case of three independent variables (his proof applies to the general case). We remark that (164) remains true for any involutive tableau whose associated constant coefficient linear P.D.E. system is (158). A special case of this result was given in Example 7.2 in Chapter IV. Returning to the general discussion, the involutivity of MA is expressed by (165)

Hk,q (MA ) = 0 for k ≥ 1.

Using this we may repeat verbatim the construction of the canonical free resolution of MA . However, in contrast to the case q = 1 the resolution does not begin with (W ∗ ⊗ SV )[q−1] (which in any case is not free if q ≥ 2), but with the free module ˜ ∗ = W ∗ ⊗ S q−1 V (= H0,0 (MA )). What is happening is that the ˜ ∗ ⊗ SV where W W homological formalism is leading us to the symbol algebra underlying the treatment of a higher order P.D.E. system as a large 1st order system. Definition 7.4. Given a tableau A ⊂ W ⊗ S q V ∗ and q ≥ 2, we define ˜ W = W ⊗ S q−1 V ∗ , and ˜⊂W ˜ ⊗V∗ A to be the image of A under the natural inclusion W ⊗ S q V ∗ ⊂ W ⊗ S q−1 V ∗ ⊗ V ∗ . Example 7.5. In the case q = 2 we let Baλij

∂ 2 ua (x) =0 ∂xi ∂xj

be the 2nd order linear homogeneous constant coefficient system corresponding to A ⊂ W ⊗ S 2 V ∗ . Then the 1st order system ∂uai (x) ∂uaj (x) − =0 ∂xj ∂xi ∂ua (x) Baλij i j = 0 ∂x corresponds to A˜ ⊂ (W ⊗ V ∗ ) ⊗ V ∗ . Returning to the general discussion, we have the following result (which would be interesting only if it were false, since it would then say that we have the wrong formalism):

332


Proposition 7.6. There is a natural isomorphism ∼

MA˜ −→ MA of graded SV -modules. This gives us another proof of the previously noted Corollary 7.7. A is an involutive tableau of order q − 1 if, and only if A˜ is an involutive tableau in the usual sense. Proof of Proposition 7.6. Using (3) and (4) above we have as subspaces of W ⊗ V ∗ ⊗ · · · ⊗ V ∗, 1 23 4 k+q

˜ ⊗ S k+1 V ∗ ∩ A˜ ⊗ S k V ∗ A˜(k) = W = W ⊗ S q−1 V ∗ ⊗ S k+1 V ∗ ∩ A ⊗ S k V ∗ where A ⊂ W ⊗ S q V ∗ . It follows from this and (155) that A˜(k) = A(k) . More precisely, A˜(k) is the image of A(k) under the natural inclusion (166)

jk : W ⊗ S q+k V ∗ ⊂ W ⊗ S q−1 V ∗ ⊗S k+1 V ∗ . ˜ ⊗ S k+1 V ∗ W

We now define the obvious degree zero graded vector space mapping µ by the diagram µ ˜ ∗ ⊗ SV − W → (W ∗ ⊗ SV )[q−1]

j ∗ W ∗ ⊗ S q−1 V ⊗ SV

where j ∗ = k≥0 jk∗ . A basic observation, whose straightforward verification we omit, is that j ∗ is induced by the multiplication S q−1 V ⊗ SV → (SV )[q−1] . This induces a commutative diagram ˜ ∗ ⊗ SV W → MA˜ → 0 ↓µ ↓ j∗ ∗ [q−1] (W ⊗ SV ) → MA → 0 where µ is now a graded module mapping and where j ∗ =

∗ k≥0 jk

with

∼

jk∗ : A˜(k)∗ −→ A(k)∗ ∼

being induced by the dual of (166). The module isomorphism j ∗ : MA˜ −→ MA is the one promised in the proposition. Finally, we shall relate the graded module MA(q) associated to the q th prolon(q) gation A ⊂ W ⊗ S q+1 V ∗ of a tableau A ⊂ W ⊗ V ∗ to MA . Denoting by + M = q≥1 Mq the positively graded sub-module of a graded module M , the simple answer is given by the


333

Proposition 7.8. There is a natural isomorphism M +(q) ∼ = (MA )[q]+ . A

Corollary 7.9. If A is involutive, then so are its prolongations A(q) . Corollary 7.10. The characteristic varieties Ξ of A and Ξ(q) of A(q) coincide. Both of these results have been proved above. Since as noted in §3 above, Corollary 7.9 is essentially a homological result, the present proof is a natural way of establishing it. We shall give a sketch of the proof in the case q = 1. The point is that we have A(1) ⊂ A ⊗ V ∗ ⊂ W ⊗ V ∗ ⊗ V ∗ ⊂ A ⊗ S k+1 V ∗ ⊂ W ⊗ 1V ∗ ⊗ ·23 · · ⊗ V ∗4

(1)(k)

A

k+2

A(k+1) ⊂ W ⊗ S k+2 V ∗ ⊂ W ⊗ V ∗ ⊗ · · · ⊗ V ∗ 1 23 4 k+2

and, as subspaces of W ⊗ V ∗ ⊗ · · · ⊗ V ∗ , 1 23 4 k+2

A(1)(k) = A(k+1) . Finally, we remark on the characteristic variety of a higher order tableau. Let A ⊂ W ⊗ V ∗(p+1) be a tableau of order p given as the kernel of a symbol mapping σ : W ⊗ V ∗(p+1) → U. For 0 = ξ ∈ V ∗ we define

σξ : W → U

by σξ (w) = σ(w ⊗ ξ ⊗ · · · ⊗ ξ ), 1 23 4 p+1

and then we define the characteristic variety of A by ΞA = {ξ ∈ PV ∗ : ker σξ = (0)}.

(167)

To justify this we consider A as an ordinary tableau A1 by the inclusion W ⊗ V ∗(p+1) ⊂ (W ⊗ V ∗ ⊗ · · · ⊗ V ∗ ) ⊗ V ∗ . 1 23 4 p ∗

∗

Setting W1 = W ⊗V ⊗· · ·⊗V we may give A1 as the kernel of a suitable mapping σ1 : W1 ⊗ V ∗ → U1 . As in the proof of (12) in §3 of Chapter V we may show that: (168)

There is a natural isomorphism ker σξ ∼ = ker σ1,ξ .

In particular, the characteristic varieties of A and A1 coincide and are given by ΞA = {ξ ∈ V ∗ : W ⊗ ξ p+1 ∩ A = 0} where ξ p+1 is the (p + 1)st symmetric product of ξ. æ

334

IX. Partial Differential Equations

CHAPTER IX PARTIAL DIFFERENTIAL EQUATIONS

In this chapter and the next, we present an introduction to the theory of overdetermined systems of partial differential equations, both linear and non-linear, as it has been developed over the last twenty five years. Rather than giving complete proofs, we have preferred in general to present many examples illustrating the various methods used in the theory. The modern theory of these systems was initially undertaken by Matsushima [1953, 1954-55] and Kuranishi [1957, 1961, 1962] within the framework of exterior differential systems. The first major result was the Cartan– Kuranishi prolongation theorem (Kuranishi [1957]). Using Ehresmann’s theory of jets, Spencer [1962] introduced fundamental new tools for the theory of overdetermined systems in order to study deformations of pseudogroup structures. In particular, to linear equations, he associated certain complexes of differential operators, namely the so-called naive and sophisticated Spencer sequences (see Example 1.13, Chapter X). Intrinsic constructions of these Spencer complexes were given by Bott [1963] and investigated by Quillen [1964]. The formal theory of overdetermined systems was then systematically studied by Goldschmidt [1967a, 1967b, 1968a, 1968b, 1970b, 1972a, 1974]; for linear equations, introductory accounts are contained in Malgrange [1966-67], Spencer [1969] and Goldschmidt [1970a]. This chapter is devoted to the basic existence theorem of Goldschmidt [1967b] for systems of non-linear partial differential equations. This result consists of two parts. First, it provides conditions which guarantee the existence of sufficiently many formal solutions for an arbitrary system. Then for an analytic system satisfying these conditions, it gives us the convergence of formal solutions and thus the existence of local solutions. We show how this theorem can be used to prove the existence of solutions for two systems involving the Ricci curvature.

§1. An Integrability Criterion. Let V and V be finite-dimensional vector spaces and let U be an open subset of Rn . Consider the system of non-linear partial differential equations of order k (1)

Φ(x, Dα u) = 0

for the unknown V -valued function u on U , where x ∈ U and Φ is a V -valued function, and α = (α1 , . . . , αn ) ranges over all multi-indices of norm ≤ k. If l ≥ 0, we say that a V -valued function u0 on a neighborhood of x0 ∈ U is an infinitesimal solution of (1) of order k + l at x0 if Dβ Φ(x, Dα u0 )|x=x0 = 0,

§1. An Integrability Criterion

335

for all multi-indices β of norm ≤ l. Clearly, the conditions for u0 to be an infinitesimal solution of order k + l at x0 are in fact imposed only on its Taylor series

p=

aα

0≤|α|≤k+l

(x − x0 )α α!

at x0 of order k + l, with aα = (Dα u0 )(x0 ), which we call a formal solution of (1) of order k + l at x0 . If m ≥ l, we say that the polynomial

q=

bβ

0≤|β|≤k+m

(x − x0 )β β!

of degree k + m extends the polynomial p if its coefficients of order ≤ k + l agree with those of p. Let Rk denote the set of all formal solutions of (1) of order k. If Φ is analytic, we are interested in finding a convergent power series solution of (1) on a neighborhood of x0 . Thus we first seek formal power series solutions of (1) at x0 . In particular, given a formal solution of (1) of order k +l, we wish to extend it to a formal solution of higher order. The aim of the existence theory of Goldschmidt [1967b] is to provide sufficient conditions under which a formal solution of order k can be extended to a formal power series solution, and in the analytic case to an analytic solution. One such condition is formal integrability: it requires that, for all l ≥ 0, every formal solution of (1) of order k + l can be extended to a formal solution of order k + l + 1. However, verifying directly the formal integrability of an equation is extremely tedious and difficult. We now present sufficient conditions for formal integrability, which in general can be effectively verified. Let u0 be an infinitesimal solution of (1) of order k at x0 and let p be the corresponding formal solution of order k at x0 . If β is a multi-index of norm k, we denote by σβ (Φ)p : V → V the derivative of Φ, considered as a function of the independent variables (x, Dα u), in the direction Dβ u at (x0 , (Dα u0 )(x0 )). We denote by Tx∗0 the cotangent space of Rn at x0 and by S m W the m-th symmetric power of a subspace W of Tx∗0 ; we write ξ m for the m-th symmetric power of an element ξ of Tx∗0 . The symbol σ(Φ)p and its first prolongation σ1 (Φ) of Φ at p are the unique linear mappings σ(Φ)p : S k Tx∗0 ⊗ V → V , σ1 (Φ)p : S k+1 Tx∗0 ⊗ V → Tx∗0 ⊗ V determined by σ(Φ)p (ξ k ⊗ v) =

(σα (Φ)p v) · ξ α ,

|α|=k

σ1 (Φ)p (ξ k+1 ⊗ v) = (k + 1)

ξ ⊗ (σα (Φ)p v) · ξ α ,

|α|=k

for ξ =

n

ξj dxj ∈ Tx∗0 and v ∈ V , where α = (α1 , . . . , αn ) and

j=1

ξ α = ξ1α1 · . . . · ξnαn .

336


We also call the kernel gk,p of the mapping σ(Φ)p the symbol of Φ at p and we set gk+1,p = Ker σ1 (Φ)p . Associated to the subspace gk,p of S k Tx∗0 ⊗ V are the Spencer cohomology groups H k+l,j (gk,p), with l, j ≥ 0, which will be defined in §2. Let {η1 , . . . , ηn } be a basis of Tx∗0 ; we say that {η1 , . . . , ηn } is a quasi-regular basis for gk at p if dim gk+1,p = dim gk,p +

n−1

dim(gk,p ∩ (S k Wj ⊗ V )),

j=1

where Wj is the subspace of Tx∗0 generated by η1 , . . . , ηj . The existence of a quasiregular basis for gk at p is equivalent to the vanishing of all these Spencer cohomology groups (see Theorem 2.14). The criterion of Goldschmidt [1967b] for the existence of formal solutions may be stated as follows. The formal solution p of order k can be extended to a formal solution of infinite order if: (i) the mapping Φ in the variables (x, Dα u), with 0 ≤ α ≤ k, is of constant rank in a neighborhood of (x0 , (Dα u0 )(x0 )); (ii) for all x in a neighborhood of x0 , there exists a formal solution of (1) of order k at x; (iii) in a neighborhood of x0 , every formal solution of (1) of order k can be extended to a formal solution of order k + 1; (iv) for all q ∈ Rk in a neighborhood of p, the rank of the linear mapping σ1 (Φ)q is independent of q; (v) for all q ∈ Rk in a neighborhood of p, the symbol gk,q at q is 2-acyclic, i.e. the cohomology groups H k+l,2 (gk,q ) vanish for l ≥ 0. The condition (v) can be replaced by the stronger condition of involutivity which need only be verified at p: (vi) there exists a quasi-regular basis of Tx∗0 for gk at p. In fact, conditions (i)–(v) imply the existence of sufficiently many formal solutions extending p, and, whenever Φ is a real-analytic function, the existence of an analytic solution u of (1) in a neighborhood of x0 satisfying (Dβ u)(x0 ) = (Dβ u0 )(x0 ), for all 0 ≤ |β| ≤ k. §2. Quasi-Linear Equations. In this section, we give an intrinsic version of the basic existence theorem of Goldschmidt [1967b]; for simplicity, here we mainly restrict our attention to quasilinear equations. We assume that all objects and mappings are differentiable of class C ∞ . In general, we do not require that the dimensions of the different components of a differentiable manifold be the same, and we allow the rank of a vector bundle over a manifold Y to vary over the different components of Y . Let X be a differentiable manifold of dimension n, whose tangent and cotangent bundles we denote by T and 5k ∗ k ∗ 6k ∗ T ∗ respectively. If k is a non-negative integer, we let T , S T and T be the k-th tensor, symmetric and exterior powers of T ∗ , respectively. We shall identify 6k ∗ 5k ∗ T with sub-bundles of T by means of the injective mappings S k T ∗ and Sk T ∗ →

5k

T ∗,

6k

T∗ →

5k

T∗

§2. Quasi-Linear Equations

337

sending the symmetric product β1 · . . . · βk , with β1 , . . . , βk ∈ T ∗ , into βσ(1) ⊗ · · · ⊗ βσ(k) , σ ∈ Sk and the exterior product β1 ∧ · · · ∧ βk into sgn σ · βσ(1) ⊗ · · · ⊗ βσ(k) , σ ∈ Sk where Sk is the group of permutations of {1, . . . , k} and sgn σ is the signature of the element σ of Sk . If Y, Z are differentiable manifolds and ρ1 : Z → X, ρ2 : Z → Y are mappings, and if F1 is a vector bundle over X and F2 is a vector bundle over Y , we denote −1 by F1 ⊗Z F2 the vector bundle ρ−1 1 F1 ⊗ ρ2 F2 over Z; if E is a vector bundle over X, we shall sometimes also denote by E the vector bundle ρ−1 1 E over Z induced by ρ1 . A fibered manifold E over X is a manifold together with a surjective submersion π : E → X. A submanifold F of E is said to be a fibered submanifold if π|F : F → X is a fibered manifold. We denote by E the sheaf of sections of E over X. Recall that two sections s and s of E over a neighborhood V of x0 ∈ X have the same k-jet at x0 if s(x0 ) = s (x0 ) and if in some, and hence in all, local coordinate systems the Taylor series of s and s agree up through order k. The class determined by s at x0 will be denoted by jk (s)(x0 ), and we write π(jk (s)(x0 )) = x0 . The set Jk (E) of all such k-jets together with the projection π is a fibered manifold over X, and x → jk (s)(x) is a section of Jk (E) over V , which we call the k-jet jk (s) of the section s; often Jk (E) is called the bundle of k-jets of sections of E. If m ≥ k and p ∈ Jm (E), we let πk (p) be the element of Jk (E) that it determines; thus πk jm (s)(x0 ) = jk (s)(x0 ). We also know that πk : Jm (E) → Jk (E) is a fibered manifold. We shall identify J0 (E) with E. Let e ∈ E with π(e) = x0 ; then there is an open neighborhood U of e and diffeomorphisms ϕ : U → Rn × Rm , ψ : πU → Rn such that the diagram ϕ

U −−−−→ Rn × Rm ⏐ ⏐ ⏐π ⏐pr1 " " ψ

πU −−−−→ Rn commutes, where pr1 is the projection onto the first factor. We obtain corresponding coordinate systems (x1 , . . . , xn , y1 , . . . , ym ) for E on U and (x1 , . . . , xn ) for X on πU . A standard local coordinate system for Jk (E) on π0−1 (U ) is (xi , yj , yαj ), where 1 ≤ i ≤ n, 1 ≤ j ≤ m and α = (α1 , . . . , αn ) ranges over all multi-indices satisfying 1 ≤ |α| ≤ k. If s is a section of E over a neighborhood of x with s(x) ∈ U , then yαj (jk (s)(x)) = Dα yj (s(x)), yj (jk (s)(x)) = yj (s(x)).

338


If E is a vector bundle, recall that Jk (E) is also a vector bundle; we have jk (s)(x0 ) + jk (s )(x0 ) = jk (s + s )(x0 ), ajk (s)(x0 ) = jk (as)(x0 ), for a ∈ R. The morphism of vector bundles : S k T ∗ ⊗ E → Jk (E) determined by (((df1 · . . . · dfk ) ⊗ s)(x)) = jk

(7

k i=1 fi

) · s (x),

where f1 , . . . , fk are real-valued functions on X vanishing at x ∈ X and s is a 7k section of E over X, is well-defined since i=1 fi vanishes to order k − 1 at x. For k < 0, we set Jk (E) = 0. One easily verifies that the sequence πk−1

0 → Sk T ∗ ⊗ E − → Jk (E) −−−→ Jk−1 (E) → 0

is exact, for k ≥ 0. Let π : E → X and π : E → X be fibered manifolds over X; a mapping ϕ : E → E is a morphism of fibered manifolds over X if π ◦ ϕ = π. We return to the study of equation (1). Let V˜ , V˜ be the trivial vector bundles U ×V , U ×V respectively. The function Φ determines a morphism ϕ : Jk (V˜ ) → V˜ of fibered manifolds over U ; in fact Φ(x, Dα u) = ϕ(jk (˜ u)(x)), where x ∈ U and u˜ is the graph of the V -valued function u on U . We identify the set Rk of formal solutions of (1) of order k with {p ∈ Jk (V˜ ) | ϕ(p) = 0}. The solutions of (1) depend only on Rk . We reserve the terminology of differential equations for such Rk satisfying the additional regularity condition that it be a fibered submanifold of the jet bundle. More precisely, we have: Definition. A (non-linear) partial differential equation Rk of order k on E is a fibered submanifold of π : Jk (E) → X. A solution s of Rk is a section of E such that jk (s) is a section of Rk . Let F be an open fibered submanifold of Jk (E) and ϕ : F → E be a morphism of fibered manifolds over X; let s be a section of E over X. We set (2)

Rk = Kers ϕ = {p ∈ F | ϕ(p) = s (π(p))}.

If (3)

s (X) ⊂ ϕ(F ),


339

the fibers of Rk are non-empty. If (3) holds and (4)

ϕ has locally constant rank,

then, according to Proposition 2.1 of Goldschmidt [1967b], Rk is a fibered submanifold of Jk (E) and a partial differential equation. The l-th prolongation of ϕ is the morphism pl (ϕ) : πk−1 F → Jl (E ) defined on the open subset πk−1 F of Jk+l (E) by pl (ϕ)(jk+l (s)(x)) = jl (ϕ ◦ jk (s))(x), for all x ∈ X and s ∈ Ex , with jk (s)(x) ∈ F . We set p0 (ϕ) = ϕ. We consider the subset Rk+l = Kerjl(s ) pl (ϕ) of πk−1 F ; the natural projection πk+l : Jk+l+1 (E) → Jk+l (E) sends Rk+l+1 into Rk+l , for all l ≥ 0. If conditions (3) and (4) hold, then Rk+l depends only on Rk and is called the l-th prolongation of Rk (see §3). We remark that any partial differential equation of order k on E can be written locally (in Jk (E)) in the form (2) with ϕ of constant rank (see Goldschmidt [1967b]). Let (x1 , . . . , xn , y1 , . . . , ym ) be the coordinate system on the open subset U of E considered above, where (x1 , . . . , xn ) is a coordinate system for X, and let (x1 , . . . , xn , z 1 , . . . , z p) be a similar coordinate system for E on an open subset U of E , with π U = πU . We consider the standard coordinate systems on the jet bundles. If ϕ(F ∩ π0−1 U ) ⊂ U , the morphism ϕ is determined by the p functions ϕr (xi , yj , yαj ),

1 ≤ |α| ≤ k,

on F ∩ π0−1 U equal to z r ◦ ϕ. The first prolongation p1 (ϕ) of ϕ is then determined by ϕr (xi , yj , yαj ) and ∂ϕr ∂ϕr i j j (x , y , y ) + (xi , yj , yαj )yl k + α ∂xk ∂yl m

l=1

1≤|β|≤k

∂ϕr i j j l (x , y , yα )yβ+k , ∂yβl

1≤l≤m

with 1 ≤ r ≤ p, 1 ≤ k ≤ n, where k is the multi-index whose k-th entry is equal to one and whose other entries are equal to zero. Example 2.1. Let F = E = Jk (E) and ϕ be the identity mapping of Jk (E). Then pl (ϕ) is the canonical imbedding λl : Jk+l (E) → Jl (Jk (E)) sending jk+l (s)(x) into jl (jk (s))(x), for x ∈ X and s ∈ Ex . We again consider equation (1); if s is the zero-section of V˜ , then Rk = Kers ϕ, and, for l ≥ 0, we may identify Rk+l with the set of formal solutions of (1) of order k + l.

340


We now return to the situation considered above. If Rk is a differential equation, we therefore call an element of Rk+l a formal solution of Rk of order k + l, and an element of R∞ = pr lim Rk+l a formal solution of Rk (of infinite order). Given a formal solution of Rk of order m ≥ k, we seek conditions which will insure that it can be extended to a formal solution. One such condition is: (5)

the mappings πk+l : Rk+l+1 → Rk+l are surjective, for all l ≥ 0.

Spencer [1962] formulated the so-called δ-Poincaré estimate, which was proved by Ehrenpreis, Guillemin and Sternberg [1965], and later by Sweeney [1967], and which gives the convergence of power series solutions for analytic partial differential equations satisfying condition (5). Malgrange [1972] (Appendix) realized that this estimate is essentially equivalent to the “privileged neighborhood theorem” of Grauert [1960] and used it together with the method of majorants to prove directly the following existence theorem for analytic differential equations. An adaptation of the proof of a result of Douady [1966] yields the required theorem of Grauert. Theorem 2.2. Suppose that X is a real-analytic manifold, that E, E are realanalytic fibered manifolds and that ϕ : F → E is a real-analytic morphism and s is an analytic section of E . Let x0 ∈ X and l ≥ 0. If πk+m : Rk+m+1,x0 → Rk+m,x0 is surjective for all m ≥ l, then given p ∈ Rk+l,x0 there exists an analytic section s of E over a neighborhood U of x0 such that jk+l (s)(x0 ) = p and jk (s)(x) ∈ Rk for all x ∈ U . If conditions (3) and (4) hold, the section s given by the theorem is a solution of the differential equation Rk . We present below sufficient conditions for (5) to hold which involve only a finite number of prolongations of Rk . We now assume that E and E are vector bundles over X. We say that the morphism of fibered manifolds ϕ : F → E is quasi-linear if there exists a morphism of vector bundles σ(ϕ) : S k T ∗ ⊗ E → E over πk−1F such that ϕ(p + u) = ϕ(p) + σ(ϕ)πk−1 p (u), for all p ∈ F , u ∈ S k T ∗ ⊗E, with p+u belonging to F . Here is the monomorphism S k T ∗ ⊗ E → Jk (E) and the vector bundles S k T ∗ ⊗ E and E are considered as induced vector bundles over πk−1F , via the mapping π. If ϕ is quasi-linear, the mapping σ(ϕ) is uniquely determined by ϕ and is called the symbol of ϕ. If ϕ is quasi-linear and σ(ϕ) is an epimorphism, and if F + (S k T ∗ ⊗ E) ⊂ F, then it is easily seen that ϕ is a surjective submersion; thus under these hypotheses, conditions (3) and (4) hold and so (2) is a differential equation.


341

Example 2.3. Let E be the vector bundle S 2 T ∗ and consider the fibered submanifold 2 ∗ S+ T of E whose sections are the positive-definite symmetric 2-forms on X. Let 2 ∗ 2 ∗ T ) of J2 (E). A section g of S+ T is a F be the open fibered submanifold J2 (S+ g Riemannian metric on X and we consider the Levi–Civita connection ∇ of g and 62 ∗ 62 ∗ T ⊗ T the Riemann curvature tensor R(g) of g, which is the section of determined by R(g)(ξ1 , ξ2 , ξ3 , ξ4 ) = g(ξ4 , (∇gξ1 ∇gξ2 − ∇gξ2 ∇gξ1 − ∇g[ξ1 ,ξ2 ] )ξ3 ), for ξ1 , ξ2 , ξ3 , ξ4 ∈ T . In fact, according to the first Bianchi identity, R(g) is a 62 ∗ 62 ∗ T ⊗ T consisting of those elements θ of section of the sub-bundle G of 62 ∗ 62 ∗ T ⊗ T which satisfy the relation θ(ξ1 , ξ2 , ξ3 , ξ4 ) + θ(ξ2 , ξ3 , ξ1 , ξ4 ) + θ(ξ3 , ξ1 , ξ2 , ξ4 ) = 0, for all ξ1 , ξ2 , ξ3 , ξ4 ∈ T ; according to Lemma 3.1 of Gasqui and Goldschmidt [1983], G is equal to the image of the morphism of vector bundles τ : S2 T ∗ ⊗ S2 T ∗ →

62

T∗ ⊗

62

T∗

defined by τ (u)(ξ1 , ξ2 , ξ3 , ξ4 ) =

1 {u(ξ1 , ξ3 , ξ2 , ξ4 ) + u(ξ2 , ξ4 , ξ1 , ξ3 ) 2 − u(ξ1 , ξ4 , ξ2 , ξ3 ) − u(ξ2 , ξ3 , ξ1 , ξ4 )},

for all u ∈ S 2 T ∗ ⊗ S 2 T ∗ and ξ1 , ξ2 , ξ3 , ξ4 ∈ T . Let E = G and let 2 ∗ T )→G Φ : J2 (S+

be the morphism of fibered manifolds over X sending j2 (g)(x) into R(g)(x), for x ∈ X. Then Φ is quasi-linear and its symbol σ(Φ) : S 2 T ∗ ⊗ S 2 T ∗ → G 2 ∗ T ) is determined by τ ; in fact, in terms of the local coordinate expression over J1 (S+ for the curvature of a metric, it is easily seen that

Φ(j2 (g)(x) + u) = R(g)(x) + τ u, 2 ∗ Tx and u ∈ (S 2 T ∗ ⊗ S 2 T ∗ )x , with x ∈ X. Since τ is an epimorphism for g ∈ S+ onto G, we see that Φ is a surjective submersion; hence if R is a section of G over X, then N2 = KerR Φ

is a differential equation, whose solutions are the Riemannian metrics g satisfying the equation R(g) = R.

342


Example 2.4. Suppose that n ≥ 3. If g is a Riemannian metric on X, we denote by Trg :

62

T∗ ⊗

62

T∗ → T∗ ⊗ T∗

the morphism defined by (Trg u)(ξ1 , ξ2 ) =

n

u(ti , ξ1 , ti , ξ2 ),

i=1

62 ∗ 62 T )x and ξ1 , ξ2 ∈ Tx , where {t1 , . . . , tn } is an orthonorfor x ∈ X, u ∈ ( T ∗ ⊗ mal basis of Tx . It is well-known that (6)

Trg (G) = S 2 T ∗

(see for example Gasqui [1982]). The Ricci curvature Ric(g) of g is the section of 2 ∗ S 2 T ∗ equal to −Trg R(g). Now as in Example 2.3, let E = S 2 T ∗ and F = J2 (S+ T ). 2 ∗ We set E = S T and let 2 ∗ ϕ : J2 (S+ T ) → S2 T ∗

be the morphism of fibered manifolds over X sending j2 (g)(x) into Ric(g)(x), for x ∈ X. Since the morphism Φ of Example 2.3 is quasi-linear, we see that this morphism ϕ is also quasi-linear and that its symbol σ(ϕ) : S 2 T ∗ ⊗ S 2 T ∗ → S 2 T ∗ 2 ∗ over J1 (S+ T ) sends (j1 (g)(x), u) into −Trg τ (u), for u ∈ (S 2 T ∗ ⊗ S 2 T ∗ )x ; in fact,

ϕ(j2 (g)(x) + u) = Ric(g)(x) − Trg τ (u), for u ∈ (S 2 T ∗ ⊗ S 2 T ∗ )x , with x ∈ X. According to (6), σ(ϕ) is an epimorphism, and so we see that ϕ is a surjective submersion. Hence if R is a section of S 2 T ∗ over X, then N2 = KerR ϕ is a differential equation, whose solutions are the Riemannian metrics g satisfying the equation (7)

Ric(g) = R.

2 ∗ Example 2.5. Suppose that n ≥ 3 and let E = S 2 T ∗ , F = J2 (S+ T ) and E = S 2 T ∗ as in Example 2.4. Let λ ∈ R and 2 ∗ T ) → S2 T ∗ ψλ : J2 (S+

be the morphism of fibered manifolds over X sending j2 (g)(x) into Ric(g)(x)−λg(x), for g ∈ S 2 Tx∗ , with x ∈ X. Since the morphism ϕ of Example 2.4 is quasi-linear, we see that ψλ is also quasi-linear and that its symbol σ(ψλ ) : S 2 T ∗ ⊗ S 2 T ∗ → S 2 T ∗


343

2 ∗ over J1 (S+ T ) is equal to σ(ϕ). Therefore ψλ is a surjective submersion; if 0 is the zero-section of S 2 T ∗ , then N2λ = Ker0 ψλ

is a differential equation whose solutions are the Riemannian metrics g satisfying the identity Ric(g) = λg, and are Einstein metrics. We now return to the situation we were considering before the above examples. Let ∆l,k : S k+l T ∗ → S l T ∗ ⊗ S k T ∗ be the natural inclusion. Let ρ : Y → X be a fibered manifold; if ψ : Sk T ∗ ⊗ E → E is a morphism of vector bundles over Y , where S k T ∗ ⊗ E and E are considered as induced vector bundles over Y via the mapping ρ, the l-th prolongation (ψ)+l : S k+l T ∗ ⊗ E → S l T ∗ ⊗ E of ψ is the morphism of vector bundles over Y equal to the composition ∆l,k ⊗ id

id ⊗ψ

S k+l T ∗ ⊗ E −−−−−→ S l T ∗ ⊗ S k T ∗ ⊗ E −−−→ S l T ∗ ⊗ E . If ϕ is quasi-linear, the l-th prolongation of σ(ϕ) (over πk−1 F ) is denoted by σl (ϕ). The following result is given by Goldschmidt [1967b], §5, and is easily verified using the standard local coordinates on the jet bundles. Proposition 2.6. If the morphism ϕ : F → E is quasi-linear, then, for l ≥ 1, so is the morphism pl (ϕ) : πk−1 F → Jl (E ) and its symbol is determined by σl (ϕ); we have (8)

pl (ϕ)(p + u) = pl (ϕ)(p) + σl (ϕ)πk−1 p (u),

for all p ∈ Jk+l (E), u ∈ S k+l T ∗ ⊗ E, with πk p ∈ F . Example 2.7. Let E, E be arbitrary vector bundles over X; assume that F = Jk (E) and that ϕ : Jk (E) → E is a morphism of vector bundles. Then D = ϕ ◦ jk : E → E is a linear differential operator of order k. In fact, any linear differential operator E → E of order k is obtained in this way. The l-th prolongation pl (ϕ) = pl (D) : Jk+l (E) → Jl (E ) of ϕ is a morphism of vector bundles over X. The symbol σ(ϕ) : S k T ∗ ⊗ E → E

344


of ϕ is the morphism ϕ ◦ of vector bundles over X. The l-th prolongation of the symbol of ϕ is the morphism of vector bundles σl (ϕ) = (σ(ϕ))+l over X. Then the diagram σl (ϕ)

S k+l T ∗ ⊗ E −−−−→ S l T ∗ ⊗ E ⏐ ⏐ ⏐ ⏐ " " pl (ϕ)

Jl (E )

−−−−→

Jk+l (E) commutes. We set p(D) = p0 (D) = ϕ,

σ(D) = σ0 (D) = σ(ϕ)

and σl (D) = σl (ϕ), for l ≥ 0. A linear differential equation of order k on E is a sub-bundle Rk of Jk (E). If ϕ has locally constant rank, then Rk = Ker ϕ is a linear differential equation, and its l-th prolongation Rk+l = Ker pl (ϕ) is a vector bundle with variable fiber which depends only on Rk . We now describe the first obstruction to the integrability of the non-linear equations determined by the morphism of fibered manifolds ϕ, which is assumed to be quasi-linear. Let W be the cokernel of the morphism σ1 (ϕ), which is a vector bundle over πk−1 F with variable fiber; if ν : T ∗ ⊗ E → W is the natural projection, the sequence (9)

σ1 (ϕ)

→ W →0 S k+1 T ∗ ⊗ E −−−→ T ∗ ⊗ E − ν

is exact. We define a mapping Ω : Rk → W as follows. If p ∈ Rk,x, with x ∈ X, let q ∈ Jk+1 (E) with πk q = p; then by (8) and the exactness of (9), the element (10)

Ω(p) = ν−1(p1 (ϕ)q − j1 (s )(x))

of Wπk−1 p is well-defined, since ϕ(p) = s (x). We denote by 0 the zero-section of W ; the following result is given by Proposition 2.1 of Goldschmidt [1972a]. Proposition 2.8. The sequence π

Ω

k −− −− −− −→ →W Rk − Rk+1 −→ 0◦π k−1

is exact, i.e. πk Rk+1 = {p ∈ Rk | Ω(p) = 0}. Proof. Let p ∈ Rk,x, with x ∈ X, and let q be an element of Jk+1 (E) satisfying πk q = p. If q ∈ Rk+1 , according to (10) we see that Ω(p) = 0. Conversely, if Ω(p) = 0, then by the exactness of (9) there exists u ∈ (S k+1 T ∗ ⊗ E)x such that σ1 (ϕ)πk−1 p u = −−1 (p1 (ϕ)q − j1 (s )(x)).


Then, by (8), we have

345

p1 (ϕ)(p + u) = j1 (s )(x)

and the element q = p + u of Jk+1 (E) belongs to Rk+1 and satisfies πk q = p. Thus Ω represents the obstruction to the surjectivity of πk : Rk+1 → Rk . The techniques for computing this first obstruction were first applied to equations arising in the theory of Lie pseudogroups by Goldschmidt [1972a, 1972b]. Subsequently, Gasqui [1975, 1979a, 1979b, 1982] studied the first obstruction and formal integrability questions for several other equations. Examples 2.4 and 2.5 (continued). Let g be a Riemannian metric on X. If 1 denotes the trivial real line bundle over X, we consider the trace mappings Tr0g : S 2 T ∗ → 1, Tr1g : T ∗ ⊗ S 2 T ∗ → T ∗ defined by Tr0g u =

n

u(ti , ti ),

i=1

(Tr1g v)(ξ) =

n

v(ti , ti , ξ),

i=1

for x ∈ X, u ∈ S 2 Tx∗ , v ∈ (T ∗ ⊗S 2 T ∗ )x , ξ ∈ Tx , where {t1 , . . . , tn } is an orthonormal basis of Tx . The Bianchi operator Bg : S 2 T ∗ → T ∗ of g is the first-order linear differential operator defined by Bg u = Tr1g ∇g u − 12 d Tr0g u, for u ∈ S 2 T ∗ . Clearly, the symbol σ(Bg ) : T ∗ ⊗ S 2 T ∗ → T ∗ of Bg is equal to Tr1g − 12 id ⊗ Tr0g . Since ∇g g = 0, we see that Bg g = 0. We recall that the Ricci curvature Ric(g) of g satisfies the Bianchi identity (11)

Bg Ric(g) = 0.

The following algebraic result is proved by Gasqui [1982] using decompositions of O(n)-modules into irreducible submodules. Lemma 2.9. The sequence of vector bundles over X (Trg ◦τ)+1

σ(Bg )

S 3 T ∗ ⊗ S 2 T ∗ −−−−−−→ T ∗ ⊗ S 2 T ∗ −−−−→ T ∗ → 0

346


is exact. Therefore the sequence of vector bundles (12)

σ1 (ϕ)

→ T∗ → 0 S 3 T ∗ ⊗ S 2 T ∗ −−−→ T ∗ ⊗ S 2 T ∗ − ν

2 ∗ 2 ∗ over J1 (S+ T ) is exact, where ν sends (j1 (g)(x), u) into σ(Bg )u, for g ∈ S+ Tx , ∗ 2 ∗ u ∈ (T ⊗ S T )x , with x ∈ X. We now compute the first obstruction to the 2 ∗ integrability of the equation N2λ . Let p = j2 (g)(x) ∈ N2λ , with x ∈ X and g ∈ S+ Tx ; then (Ric(g) − λg)(x) = 0. By (10), (11) and the exactness of (12), we have

Ω(p) = ν−1 j1 (Ric(g) − λg)(x) = σ(Bg )−1 j1 (Ric(g) − λg)(x) = Bg (Ric(g) − λg)(x) = 0, since Bg g = 0. Therefore if N3λ denotes the first prolongation of N2λ , the mapping π2 : N3λ → N2λ is surjective, by Proposition 2.8. The computation of the first obstruction Ω for the equation N2 of Example 2.4 2 ∗ is quite similar. Namely, if p = j2 (g)(x) ∈ N2 , with x ∈ X and g ∈ S+ Tx , then (Ric(g) − R)(x) = 0 and Ω(p) = ν−1j1 (Ric(g) − R)(x) = σ(Bg )−1 j1 (Ric(g) − R)(x) = Bg (Ric(g) − R)(x) = −(Bg R)(x). Thus by Proposition 2.8, we see that p ∈ π2 N3 if and only if (Bg R)(x) = 0. For a general section R, the mapping π2 : N3 → N2 is not surjective (see DeTurck [1981, 1982]). In fact, a solution g of (7) is also a solution of the equation Bg R = 0, and this will be taken into account in the next example. æ Example 2.10. If h ∈ S 2 T ∗ , let h : T → T ∗ be the mapping determined by h(ξ, η) = η, h (ξ), for ξ, η ∈ T . If h is non-degenerate, that is, if h is an isomorphism, we denote by h : T ∗ → T the inverse of h . We suppose that n ≥ 3 and consider the objects of Example 2.4. The proofs of the following algebraic lemmas due to DeTurck [1981] can be proved using the methods of Gasqui [1982] involving decompositions into irreducible O(n)modules; here g denotes a Riemannian metric on X.


347

Lemma 2.11. The mapping (Trg ◦τ)⊕σ1 (Bg )

S 2 T ∗ ⊗ S 2 T ∗ −−−−−−−−−−→ S 2 T ∗ ⊕ (T ∗ ⊗ T ∗ ) is an epimorphism of vector bundles. Lemma 2.12. The sequence of vector bundles (Trg ◦τ)+1 ⊕σ2 (Bg )

µg

S 3 T ∗ ⊗ S 2 T ∗ −−−−−−−−−−−→ (T ∗ ⊗ S 2 T ∗ ) ⊕ (S 2 T ∗ ⊗ T ∗ ) −→ T ∗ → 0 is exact, where µg sends u⊕v, with u ∈ T ∗ ⊗S 2 T ∗ and v ∈ S 2 T ∗ ⊗T ∗ , into σ(Bg )u. Let h be a section of S 2 T ∗ . We define a section Lg (h) of S 2 T ∗ ⊗ T by 1 {(∇gξ h)(η, ζ) + (∇gη h)(ζ, ξ) − (∇gζ h)(ξ, η)}, 2 5k ∗ 5k+1 ∗ for ξ, η, ζ ∈ T . If ω ∈ T , let Lg (h)ω be the element of T given by (13)

(14)

g(Lg (h)(ξ, η), ζ) =

(L (h)ω)(ξ, ξ1 , . . . , ξk ) = − g

k

ω(ξ1 , . . . , ξj−1, Lg (h)(ξ, ξj ), ξj+1 , . . . , ξk ),

j=1

for ξ, ξ1 , . . . , ξk ∈ T . By means of (14), it is easily seen that Lg (h)R is a section of T ∗ ⊗ S 2 T ∗ satisfying the relation (15)

ξ, σ(Bg )(Lg (h)R) = −R((Tr0g ⊗ id)Lg (h), ξ),

for ξ ∈ T . From (13), it follows directly that (16)

Bg (h) = g (Tr0g ⊗ id)Lg (h).

If R is a non-degenerate section of S 2 T ∗ , from (15) and (16) we deduce that (17)

Bg (h) = − g · R (σ(Bg )(Lg (h)R)).

Let x ∈ X and assume that h(x) = 0; then g + h is a Riemannian metric on a neighborhood U of x and ∇g+h − ∇g : T → T ∗ ⊗ T is a differential operator of order zero on U which arises from a section of S 2 T ∗ ⊗ T . In fact, we have ∇g+h − ∇g = Lg (h) at x; the verification of this relation is essentially the same as that of identity (4.8) of Gasqui and Goldschmidt [1983]. Thus (∇g+h R)(x) = (∇g R)(x) + (Lg (h)R)(x), and so (18)

Bg+h (R) = Bg (R) + σ(Bg )(Lg (h)R)

348


at x. We now assume that R is a non-degenerate section of S 2 T ∗ . We consider the morphism of fibered manifolds 2 ∗ ψR : J1 (S+ T ) → T∗

defined by ψR (j1 (g)(x)) = (g · R )(Bg (R))(x). According to (18) and (17), we have ψR (j1 (g + h)(x)) = ψR (j1 (g)(x)) − Bg (h)(x) = ψR (j1 (g)(x)) − σ(Bg )−1 j1 (h)(x). Thus ψR is quasi-linear and its symbol σ(ψR ) : T ∗ ⊗ S 2 T ∗ → T ∗ 2 ∗ over S+ T sends (g(x), u) into −σ(Bg )u, for u ∈ (T ∗ ⊗ S 2 T ∗ )x , and is surjective by Lemma 2.9. Therefore, according to Proposition 2.6, the morphism of fibered manifolds 2 ∗ Ψ = ϕ ⊕ p1 (ψR ) : J2 (S+ T ) → S 2 T ∗ ⊕ J1 (T ∗ )

is quasi-linear, and its symbol σ(Ψ) : S 2 T ∗ ⊗ S 2 T ∗ → S 2 T ∗ ⊕ J1 (T ∗ ) 2 ∗ over J1 (S+ T ) at p = j1 (g)(x) is determined by

−{(Trg ◦ τ ) ⊕ σ1 (Bg )} : (S 2 T ∗ ⊗ S 2 T ∗ )x → (S 2 T ∗ ⊕ (T ∗ ⊗ T ∗ ))x ; in fact, σ(Ψ)(j1 (g)(x), u) = −(Trg τ (u) ⊕ σ1 (Bg )u), 2 ∗ for u ∈ (S 2 T ∗ ⊗ S 2 T ∗ )x . Let x ∈ X and g0 ∈ S+ Tx . Since the symbol of ψR 2 ∗ is surjective, there exists p ∈ J1 (S+ T )x with ψR (p) = 0 and π0 (p) = g0 . By 2 ∗ T ) induced Lemma 2.11, the image of σ(Ψ) is equal to the vector bundle over J1 (S+ 2 ∗ ∗ ∗ 2 ∗ T ) from S T ⊕ (T ⊗ T ); hence it is easily seen that there exists q ∈ J2 (S+ satisfying π1 (q) = p and Ψ(q) = R(x) ⊕ 0 and that Ψ is a submersion. Therefore

N2 = KerR⊕0 Ψ is a differential equation of order 2 satisfying 2 ∗ π0 N2 = S+ T .

(19)

Its solutions are the same as of those of equation (7). The diagram 2 ∗ J3 (S+ T ) 8 ⏐ ⏐id

−−−−→

p1 (Ψ)

J1 (S 2 T ∗ ) ⊕ J1 (J1 (T ∗ )) 8 ⏐ ⏐id ⊕λ1

p1 (ϕ)⊕p2 (ψR )

J1 (S 2 T ∗ ) ⊕ J2 (T ∗ )

2 ∗ J3 (S+ T ) −−−−−−−−−→


349

is commutative, where λ1 is injective; hence the first prolongation N3 of N2 is equal to Kerj1 (R)⊕0 p1 (ϕ) ⊕ p2 (ψR ). By Proposition 2.6, we have (p1 (ϕ) ⊕ p2 (ψR ))(p + u) = (p1 (ϕ) ⊕ p2 (ψR ))(p) − (((Trg ◦ τ )+1 u) ⊕ (σ2 (Bg )u)), 2 ∗ T ), u ∈ (S 3 T ∗ ⊗ S 2 T ∗ )x . Therefore by Lemma 2.12, the for p = j3 (g)(x) ∈ J3 (S+ first obstruction Ω : N2 → T ∗

for the equation N2 is well-defined by (20)

Ω (p) = µg (−1 j1 (Ric(g) − R)(x) ⊕ −1 j2 (g · R (Bg (R)))(x)),

2 ∗ for p = j2 (g)(x) ∈ N2 , with x ∈ X and g ∈ S+ Tx satisfying

(Ric(g) − R)(x) = 0 and j1 (g · R (Bg (R)))(x) = 0.

(21)

It is easily verified that the sequence 2 N3 −→ N2

π

Ω −→ −→ 0◦π

T∗

is exact. Moreover, by (20), (11) and (21), we have Ω (p) = σ(Bg )−1 j1 (Ric(g) − R)(x) = (Bg (Ric(g) − R))(x) = −(Bg R)(x) = 0. Thus, the mapping π2 : N3 → N2 is surjective. For k ≥ 0, we denote by δ = ∆1,k : S k+1 T ∗ → T ∗ ⊗ S k T ∗ the natural inclusion; we have δ(β1 · . . . · βk+1 ) =

k+1

βi ⊗ β1 · . . . · βî · . . . · βk+1 ,

i=1

for all β1 , . . . , βk+1 ∈ T ∗ , where the symbol ˆ above a letter means that it is omitted. We extend δ to a morphism of vector bundles δ:

6j

T ∗ ⊗ S k+1 T ∗ →

6j+1

T ∗ ⊗ Sk T ∗

350


6j ∗ sending ω ⊗ u into (−1)j ω ∧ δu, for all ω ∈ T and u ∈ S k+1 T ∗ . If we set l ∗ S T = 0, for l < 0, the Poincaré lemma for forms with polynomial coefficients implies that the sequence → T ∗ ⊗ S k−1 T ∗ − → 0 → Sk T ∗ − δ

(22)

δ

62

T ∗ ⊗S k−2 T ∗ − → ··· 6n ∗ T ⊗ S k−n T ∗ → 0 → δ

is exact, for k ≥ 1. Let ρ : Y → X be a fibered manifold and let ψ : Sk T ∗ ⊗ E → E be a morphism of vector bundles over Y , where S k T ∗ ⊗ E and E are considered as induced vector bundles over Y via the mapping ρ. Let gk be the kernel of ψ, which is a vector bundle over Y with variable fiber. For l ≥ 0, the kernel gk+l of the l-th prolongation (ψ)+l of ψ is equal to (S k+l T ∗ ⊗ E) ∩ (S l T ∗ ⊗ gk ) and is called the l-th prolongation of gk . We set gk+l = S k+l T ∗ ⊗ E, considered as a vector bundle over Y , for l < 0. It is easily seen that the diagram 6j (23)

id ⊗(ψ)+(l+1)

T ∗ ⊗ S k+l+1 T ∗ ⊗ E −−−−−−−−→ ⏐ ⏐ "δ

6j+1

id ⊗(ψ)+l

T ∗ ⊗ S k+l T ∗ ⊗ E −−−−−−−−→

6j

T ∗ ⊗ S l+1 T ∗ ⊗ E ⏐ ⏐ "δ

6j+1

T ∗ ⊗ Sl T ∗ ⊗ E

commutes, and so the morphism δ induces by restriction mappings δ:

6j

T ∗ ⊗ gk+l+1 →

6j+1

T ∗ ⊗ gk+l ;

the cohomology of the complexes (24)

→ T ∗ ⊗ gm−1 − → 0 → gm − δ

δ

62

T ∗ ⊗ gm−2 − → ··· → δ

6n

T ∗ ⊗ gm−n → 0

is the Spencer cohomology of gk . We denote by H m−j,j (gk ) the cohomology of (24) 6 at j T ∗ ⊗ gm−j . We say that gk is r-acyclic if H m,j (gk ) = 0, for all m ≥ k and 0 ≤ j ≤ r, and that gk is involutive if it is n-acyclic. It is easily seen that gk is always 1-acyclic. The following theorem asserts that all but a finite number of these cohomology groups vanish, whenever there is an integer d such that dim Ex ≤ d for all x ∈ X (see Quillen [1964], Sweeney [1968]). Theorem 2.13. If there is an integer d for which dim Ex ≤ d, for all x ∈ X, then there exists an integer k0 depending only on n, k and d such that H m,j (gk ) = 0,


351

for all m ≥ k0 , j ≥ 0. Let x ∈ X and {t1 , . . . , tn } be a basis of Tx . If {α1 , . . . , αn } is the basis of Tx∗ dual ∗ to {t1 , . . . , tn }, then we denote by S k Tx,{t the subspace of S k Tx∗ generated 1 ,...,tj } by the symmetric products αi1 · . . . · αik , with j + 1 ≤ i1 ≤ · · · ≤ ik ≤ n. If y ∈ Y , with ρ(y) = x, we set ∗ gk,y,{t1,...,tj } = gk,y ∩ (S k Tx,{t ⊗ Ex ). 1 ,...,tj }

We say that {t1 , . . . , tn } is a quasi-regular basis for gk at y if dim gk+1,y = dim gk,y +

n−1

dim gk,y,{t1,...,tj } .

j=1

The following criterion for the involutivity of gk is due to Serre (see Guillemin and Sternberg [1964], Appendix; see also §§2,3, Chapter VIII). Theorem 2.14. The following conditions are equivalent: (i) there exists a quasi-regular basis of Tx for gk at y; (ii) H m,j (gk )y = 0, for all m ≥ k, j ≥ 0. ´ Cartan’s proof of An elementary argument, due to Sternberg and based on E. the Poincaré lemma, shows that condition (i) of the above theorem implies (ii). Any basis of Tx is quasi-regular for the sub-bundle T ∗ of S 1 T ∗ ; since S k+1 T ∗ is its k-th prolongation, this argument proves that the sequence (22) is exact (see §2, Chapter VIII). Using the preceding theorem, it is easily seen that: Lemma 2.15. If gk+1 is a vector bundle over Y and gk is involutive at y0 ∈ Y , then gk is involutive for all y in a neighborhood of y0 . We again consider the morphism ϕ : F → E , where F is an open fibered submanifold of Jk (E). Assume that ϕ is quasi-linear and suppose that conditions (3) and (4) hold; then Rk is a differential equation. The symbol gk of Rk and the l-th prolongation gk+l of the symbol of Rk are the vector bundles with variable fiber over Rk whose fibers at p ∈ Rk are (25) (26)

gk,p = Ker σ(ϕ)πk−1 p , gk+l,p = Ker σl (ϕ)πk−1 p = (gk )+l,p ;

they depend only on Rk and not on ϕ (see §3). We say that the differential equation Rk is formally integrable if: (i) gk+l+1 is a vector bundle over Rk , for all l ≥ 0; (ii) πk+l : Rk+l+1 → Rk+l is surjective, for all l ≥ 0. The following result (Goldschmidt [1967b], Theorem 8.1) is the formal part of the basic existence theorem for quasi-linear morphisms; it gives us a criterion for formal integrability. Together with Theorem 2.2, it provides us with the existence of analytic solutions for analytic quasi-linear partial differential equations.

352


Theorem 2.16. Assume that (3) and (4) hold. If (A) gk+1 is a vector bundle over Rk , (B) πk : Rk+1 → Rk is surjective, (C) gk is 2-acyclic, then Rk is formally integrable. In fact, if (3), (4), (A) and (B) hold, Goldschmidt [1967b] constructs the second obstruction to the integrability of Rk , which is a mapping κ : Rk+1 → H k,2 (gk ) over Rk , and then proves that the sequence πk+1

Rk+2 −−−→ Rk+1

κ − −− −→ → 0◦πk

H k,2 (gk )

is exact. Thus the higher obstructions to integrability lie in the cohomology groups H k+l,2 (gk ) and condition (C) then implies that the mappings πk+l : Rk+l+1 → Rk+l are surjective for l > 0. A complete proof of Theorem 2.16 for linear equations will be given in Chapter X (Theorem 1.6); in particular, in the course of this proof, the mapping κ will be constructed. According to Theorem 2.14, we may replace condition (C) in the above theorem by: (C ) for all p ∈ Rk , there exists a quasi-regular basis of Tπ(p) for gk at p. We remark that conditions (A), (B) and (C ) are of “finite type”, in the sense that they involve only ϕ and σ(ϕ) and their first prolongations. From Theorems 2.13 and 2.16, Goldschmidt [1967b] deduces the following version of the Cartan–Kuranishi prolongation theorem (Kuranishi [1957]), which asserts that the condition of formal integrability is of “finite type”: to determine whether Rk is formally integrable, we need examine only a finite number of prolongations of ϕ. Theorem 2.17. Assume that X is connected and that (3) and (4) hold. There exists an integer k0 ≥ k depending only on n, k and the rank of E such that, if (i) gk+l+1 is a vector bundle over Rk , for all 0 ≤ l ≤ k0 − k, (ii) πk+l : Rk+l+1 → Rk+l is surjective, for all 0 ≤ l ≤ k0 − k, then Rk is formally integrable. We now show how Theorems 2.16 and 2.2 give us the existence of solutions for the equations of Examples 2.5 and 2.10 in the analytic case. Examples 2.5 and 2.10 (continued). For the differential equations N2λ and N2 we have verified condition (B). The involutivity of the symbols of N2λ and N2 will be a consequence of the following lemma of Gasqui [1982] and DeTurck [1981]. Lemma 2.18. Let g be a Riemannian metric on X and x ∈ X. An orthonormal basis of Tx is quasi-regular for the kernels of the morphisms (27)

Trg ◦ τ : S 2 T ∗ ⊗ S 2 T ∗ → S 2 T ∗ ,

(28)

(Trg ◦ τ ) ⊕ σ1 (Bg ) : S 2 T ∗ ⊗ S 2 T ∗ → S 2 T ∗ ⊕ (T ∗ ⊗ T ∗ )

§3. Existence Theorems

353

at x. We remark that the first prolongations of the kernels of the morphisms (27) and (28) are equal to the kernels of (Trg ◦ τ )+1 and (Trg ◦ τ )+1 ⊕ σ2 (Bg ) respectively. Let g be a Riemannian metric on X and x ∈ X; set p = j2 (g)(x). If p ∈ N2λ (resp. N2 ), then the fiber of the symbol of N2λ (resp. N2 ) is the kernel of (27) (resp. (28)) at x. Thus by Lemma 2.18 condition (C ) holds for N2λ and N2 , while condition (A) for these equations is a consequence of Lemmas 2.9 and 2.12. Therefore, by Theorem 2.16, N2λ and N2 are formally integrable. Assume that X is a real-analytic manifold. Then ψλ and N2λ are analytic. If R is an analytic non-degenerate section of S 2 T ∗ , then Ψ and N2 are analytic. The following result of DeTurck [1981] is now a direct consequence of Theorem 2.2 and (19); its proof outlined here is a variant of the one given by DeTurck. Theorem 2.19. Let X be a real-analytic manifold of dimension n ≥ 3 and let R 2 ∗ Tx , there be an analytic non-degenerate section of S 2 T ∗ . If x ∈ X and g0 ∈ S+ exists an analytic Riemannian metric g on a neighborhood of x such that g(x) = g0 ,

Ric(g) = R.

The following theorem is due to Gasqui [1982]. Theorem 2.20. Let X be a real-analytic manifold of dimension n ≥ 3 and x ∈ X. Let g0 be a Riemannian metric on X and R0 ∈ Gx such that −Trg0 R0 = λg0 (x), with λ ∈ R. Then there exists an analytic Riemannian metric g on a neighborhood of x such that g(x) = g0 , R(g)(x) = R0 , Ric(g) = λg. Proof. Since the morphism Φ of Example 2.3 is quasi-linear and its symbol is sur2 ∗ jective, we see that there is an element p of J2 (S+ T )x satisfying π0 (p) = g0 (x) and Φ(p) = R0 . From our hypothesis on R0 , we see that p ∈ N2λ . Because N2λ is formally integrable, Theorem 2.2 gives us an analytic solution g of N2λ over a neighborhood of x such that j2 (g)(x) = p. §3. Existence Theorems. We now briefly show how the results of §2 can be generalized to arbitrary systems of partial differential equations. As the equations are in general no longer quasilinear, we must consider the structure of affine bundle which the jet bundles possess. We no longer assume that E and E are vector bundles, but continue to suppose that they are fibered manifolds over X. We denote by V (E) the bundle of vectors tangent to the fibers of π : E → X. According to Proposition 5.1 of Goldschmidt [1967b], for k ≥ 1 the jet bundle Jk (E) is an affine bundle over Jk−1 (E) modeled on the vector bundle S k T ∗ ⊗Jk−1 (E) V (E) over Jk−1(E). In fact, if p ∈ Jk−1 (E) with π(p) = x, the vector space S k Tx∗ ⊗ Vπ0 p (E) considered as an additive group acts freely and transitively on the fiber of Jk (E) over p; for u ∈ S k Tx∗ ⊗ Vπ0 p (E), we denote by u + q the image

354


of the element q of Jk (E)p under the action of u. In terms of the standard local coordinate system for Jk (E) considered in §2, this action can be described as follows. If α = (α1 , . . . , αn) is a multi-index of norm k, we consider the section dxα = (dx1 )α1 · . . . · (dxn )αn of S k T ∗ . If an element q of Jk (E) satisfies π0 q ∈ U and π(q) = x, and if (xi , yj , yαj ) are the coordinates of q and

u=

ajα (dxα )(x) ⊗

|α|=k

∂ (π0 q), ∂yj

the coordinates of u + q are (xi , yj , zαj ), where zαj =

yαj + ajα , if |α| = k, if 1 ≤ |α| < k. yαj ,

An intrinsic definition of this action is given by Goldschmidt [1967b], §5. If E is a vector bundle, then Vπ0 p (E) is canonically isomorphic to Ex and so S k Tx∗ ⊗ Vπ0 p (E) S k T ∗ ⊗ Ex; in this case, the action of S k Tx∗ ⊗ Vπ0 p (E) on Jk (E) can be described in terms of the vector bundle structure of Jk (E) and the morphism : u + q = (u) + q, where on the left-hand side u is considered as an element of S k Tx∗ ⊗ Vπ0 p (E), while on the right-hand side u is viewed as an element of (S k T ∗ ⊗ E)x . From the affine bundle structure of Jk (E), we obtain a morphism of vector bundles µ : S k T ∗ ⊗Jk (E) V (E) → V (Jk (E)), ∗ ⊗ Vπ0 p (E), into the tangent vector sending (p, u), with p ∈ Jk (E) and u ∈ S k Tπ(p)

d (p + tu)|t=0 , dt where t ∈ R. It is easily seen that the sequence µ

πk−1∗

−1 0 → S k T ∗ ⊗Jk (E) V (E) − → V (Jk (E)) −−−→ πk−1 V (Jk−1 (E)) → 0

of vector bundles over Jk (E) is exact (see Goldschmidt [1967b], Proposition 5.2). We shall identify S k T ∗ ⊗Jk (E) V (E) with its image in V (Jk (E)) under the mapping µ. Let F be an open fibered submanifold of Jk (E) and let ϕ : F → E be a morphism of fibered manifolds over X. The mapping ϕ∗ : V (F ) → V (E ) induces a morphism ϕ∗ = ϕ∗ ◦ µ : S k T ∗ ⊗F V (E) → V (E ),


355

which we also denote by σ(ϕ); let σl (ϕ) be the composition ∆l,k ⊗ id

id ⊗ϕ∗

S k+l T ∗ ⊗F V (E) −−−−→ S l T ∗ ⊗ S k T ∗ ⊗F V (E) −−−→ S l T ∗ ⊗E V (E ). According to Proposition 5.6 of Goldschmidt [1967b], for l ≥ 1 the mapping pl (ϕ) : πk−1 F → Jl (E ) is a morphism of affine bundles over pl−1 (ϕ) whose associated morphism of vector bundles is induced by σl (ϕ); in other words, (29)

pl (ϕ)(u + p) = σl (ϕ)πk p (u) + pl (ϕ)p,

∗ for all p ∈ Jk+l (E), with πk p ∈ F , and u ∈ S k+l Tπ(p) ⊗ Vπ0 p (E). This formula is easily verified using the standard local coordinates on jet bundles. If E, E are vector bundles and ϕ is quasi-linear, then σl (ϕ) can be identified with the mapping of §2 denoted there by σl (ϕ); moreover, Proposition 2.6 can be deduced from formula (29). Let Rk be a differential equation of order k on E. The l-th prolongation of Rk is the subset Rk+l of Jk+l (E) determined by the equality

λl Rk+l = Jl (Rk ) ∩ λl Jk+l (E), where Jl (Rk ) is considered as a subset of Jl (Jk (E)). The projection πk+l : Jk+l+1 (E) → Jk+l (E) sends Rk+l+1 into Rk+l . The symbol of Rk is the sub-bundle with varying fiber gk = V (Rk ) ∩ (S k T ∗ ⊗Rk V (E)) of (S k T ∗ ⊗Jk (E) V (E))|Rk , which is equal to the kernel of the morphism of vector bundles (S k T ∗ ⊗Jk (E) V (E))|Rk → (V (Jk (E))|Rk )/V (Rk ) over Rk . Let gk+l be the l-th prolongation of gk ; it is a sub-bundle with varying fiber of S k+l T ∗ ⊗Rk V (E). If s is a section of E over X and Rk is given by (2), and if conditions (3) and (4) are satisfied, then the l-th prolongation Rk+l of Rk is equal to Kerjl (s ) pl (ϕ) and gk = (Ker σ(ϕ))|Rk , gk+l = (Ker σl (ϕ))|Rk ; thus in this case, our definition of Rk+l coincides with the one given in §2. Using the identity (29), the first obstruction Ω : Rk → W to the integrability of Rk , where W is a vector bundle with variable fiber over Rk , can be constructed in a way similar to that of §2; then we still have πk Rk+1 = {p ∈ Rk | Ω(p) = 0} (see Goldschmidt [1972a], Proposition 2.1). If moreover E, E are vector bundles and ϕ is quasi-linear, the morphisms σ(ϕ) and σl (ϕ) can be identified with the mappings of §2, while gk and gk+l are given by (25) and (26). We say that a differential equation Rk of order k on E is formally integrable if: (i) gk+l+1 is a vector bundle over Rk , for all l ≥ 0;

356


(ii) πk+l : Rk+l+1 → Rk+l is surjective, for all l ≥ 0. If Rk is formally integrable, then for l ≥ 0, according to Proposition 7.2 of Goldschmidt [1967b], Rk+l is a fibered submanifold of Jk+l (E) and πk+l : Rk+l+1 → Rk+l is an affine sub-bundle of πk+l : Jk+l+1 (E)|Rk+l → Rk+l , whose associated vector bundle is the vector bundle πk−1 gk+l+1 over Rk+l induced from gk+l+1 by πk : Rk+l → Rk ; this last statement implies that for each p ∈ Rk+l the fiber of Rk+l+1 over p is an affine subspace of the fiber of Jk+l+1 (E) over p whose associated vector space is gk+l+1,πk p . Theorem 2.16 can be viewed as a special case of Theorem 8.1 of Goldschmidt [1967b], which we now state as Theorem 3.1. Let Rk be a differential equation of order k on E. If (A) gk+1 is a vector bundle over Rk , (B) πk : Rk+1 → Rk is surjective, (C) gk is 2-acyclic, then Rk is formally integrable. Again as for Theorem 2.16, according to Theorem 2.14, we may replace condition (C) in the above theorem by: (C ) for all p ∈ Rk , there exists a quasi-regular basis of Tπ(p) for gk at p. Theorem 2.17 also holds for a differential equation of order k on E, with k0 ≥ k depending only on n, k and the dimension of E. Theorem 2.2 now provides us with the existence of analytic solutions: Theorem 3.2. Suppose that X is a real-analytic manifold and that E is a realanalytic fibered manifold. If Rk is an analytic formally integrable differential equation of order k on E, then for all p ∈ Rk+l there exists an analytic solution s of Rk on a neighborhood of x = π(p) such that jk+l (s)(x) = p. We now present the intrinsic formulation of the criterion of §1 for the existence of analytic solutions of analytic equations; it is a direct consequence of Theorems 3.1 and 2.2. Theorem 3.3. Assume that X is a real-analytic manifold and that E, E are realanalytic fibered manifolds. Let F be an open fibered submanifold of Jk (E) and let ϕ : F → E be an analytic morphism of fibered manifolds over X and s be an analytic section of E over X. Let Rk = Kers ϕ, gk = (Ker σ(ϕ))|Rk ,

Rk+1 = Kerj1 (s ) p1 (ϕ), gk+1 = (Ker σ1 (ϕ))|Rk ,

and let p ∈ Rk . If there exists a neighborhood U of p in F such that: (i) ϕ has constant rank on U , (ii) s (πU ) ⊂ ϕ(F ), (iii) πk : Rk+1 ∩ πk−1 (U ) → Rk ∩ U is surjective, (iv) gk+1 |(U ∩Rk ) is a vector bundle, (v) H k+l,2 (gk )|(U ∩Rk ) = 0, for all l ≥ 0, then there exists an analytic section s of E over a neighborhood V of x0 = π(p) such that jk (s)(x0 ) = p and jk (s)(x) ∈ Rk for all x ∈ V . According to Lemma 2.15, the condition (v) can be replaced by the stronger condition:


357

(vi) there exists a quasi-regular basis of Tx0 for gk at p. Kuranishi [1967] proves the existence of analytic solutions of analytic partial differential equations by the method of Cartan–K¨ ahler, using the Cauchy–Kowalewski theorem, under these assumptions (i)–(iv) and (vi). æ

358

X. Linear Differential Operators

CHAPTER X LINEAR DIFFERENTIAL OPERATORS

In this chapter, we consider only linear systems of partial differential equations, and use the notation and terminology introduced in Chapter IX. In general, if D : E → F is a linear differential operator, where E, F are vector bundles over the manifold X, and if f is a section of F , the inhomogeneous equation Du = f is not solvable for a section u of E unless f satisfies a requisite compatibility condition. Indeed, certain conditions must be imposed on the formal power series expansion j∞ (f)(x) of f at x ∈ X in order that it may be written as j∞ (Du)(x), for some section u of E. Under certain regularity assumptions on D, they can be expressed in terms of a differential operator P : F → B of finite order, where B is a vector bundle over X. This operator is called the compatibility condition for D and is obtained by repeatedly differentiating the equation. We then obtain a complex of differential operators D

P

E −→ F −→ B which is exact at the formal power series level: the formal power series expansion j∞ (f)(x) of f at x can be written in the form j∞ (Du)(x), for some section u of E, if and only if P f vanishes to infinite order at x. For example, the inhomogeneous equation du = f, where u is a real-valued function and f is a 1-form on X, is not solvable for u unless df = 0. In Section 1, we present a complete proof of the formal existence theorem of Goldschmidt [1967a] for homogeneous linear systems (Theorem 1.6), and existence results for the compatibility condition of a linear differential operator, as well as existence theorems for analytic differential operators. We also construct formally exact complexes of differential operators, whose vector bundles can be explicitly described under specific hypotheses; these include the sophisticated Spencer sequence (see Theorems 1.9 and 1.11, Examples 1.10, 1.12 and 1.13). Section 2 is devoted to various examples of these complexes, while Section 3 is concerned with exactness results for our complexes under the additional assumption of ellipticity. §1. Formal Theory and Complexes. Let E, F be vector bundles over the differentiable manifold X of dimension n. We denote by C ∞(U, E) the space of sections of the vector bundle E over an open subset U of X. Let ϕ : Jk (E) → F be a morphism of vector bundles and let D = ϕ ◦ jk : E → F be the corresponding differential operator of order k. We consider the morphisms associated to ϕ and D in Example 2.7 of Chapter IX. Let J∞ (E) = pr lim Jm (E)

§1. Formal Theory and Complexes

359

be the bundle of jets of infinite order of sections of E, and set p∞ (D) = pr lim pm (D). If k = 1, then D(fs) = σ(D)(df ⊗ s) + fDs,

(1)

where f is a real-valued function on X and s is a section of E over X. Indeed, this formula holds if f is a constant function; on the other hand, if x ∈ X and f(x) = 0, then it is valid at x, according to the definitions of and σ(D). Let Rk be the kernel of ϕ and Rk+l the kernel of pl (ϕ). The symbol gk of (Rk , ϕ) is the kernel of σ(ϕ) : S k T ∗ ⊗ E → F and its l-th prolongation gk+l is the kernel of the morphism of vector bundles σl (ϕ) = (σ(ϕ))+l : S k+l T ∗ ⊗ E → S l T ∗ ⊗ F over X. For l < 0, we write Rk+l = Jk+l (E) and gk+l = S k+l T ∗ ⊗ E; then the sequence πk+l−1 0 → gk+l − → Rk+l −−−−→ Rk+l−1 is exact. We recall that, if ϕ has locally constant rank, Rk is a linear differential equation and that the l-th prolongation Rk+l of Rk is determined by the equality λl Rk+l = Jl (Rk ) ∩ λl Jk+l (E). If Rk is a vector bundle, we call gk the symbol of the equation Rk . If Rk+l is a vector bundle, it is easily verified that the diagram pl+m (ϕ)

0 −−−−→ Rk+l+m −−−−→ Jk+l+m (E) −−−−−→ ⏐ ⏐ "id

Jl+m (F ) ⏐ ⏐ "λm

pm (pl (ϕ))

0 −−−−→ (Rk+l )+m −−−−→ Jk+l+m (E) −−−−−−→ Jm (Jl (F )) is commutative and exact, where λm is injective; therefore the m-th prolongation (Rk+l )+m of the equation Rk+l is equal to Rk+l+m . It follows that the m-th prolongation (gk+l )+m of gk+l is equal to gk+l+m . We say that Rk is formally integrable if: (i) Rk+l is a vector bundle for all l ≥ 0; (ii) πk+l : Rk+l+1 → Rk+l is surjective for all l ≥ 0. We recall the following lemma of Goldschmidt [1967a] which we will require later. Lemma 1.1. If ψ

ψ

E −→ E −−→ E is an exact sequence of vector bundles over X, then the kernel of ψ and the cokernel of ψ are both vector bundles. Let B be a vector bundle over X. If D : F → B is a differential operator of order l, we say that the sequence (2)

D

D

E −→ F −→ B

360


is formally exact if the sequence p∞ (D )

p∞ (D)

J∞ (E) −−−−→ J∞ (F ) −−−−−→ J∞ (B) is exact. Example 1.2. Suppose that X = Rn and that (2) is a complex of constant coefficient differential operators. If U is an open convex subset of Rn , according to the Ehrenpreis–Malgrange theorem (see Ehrenpreis [1970], Malgrange [1963] and Hörmander [1973], §7.6), the sequence D

C ∞ (U, E) −→ C ∞(U, F ) −→ C ∞ (U, B) D

is exact. Unfortunately, formal exactness is not a good concept for operators with variable coefficients, even for analytic operators, as we shall see below with Example 1.5. It shall be replaced by stronger conditions (see Theorem 1.4). Lemma 1.3. Let ϕ : Jk (E) → F , ψ : Jl (F ) → B be morphisms of vector bundles over X, and set D = ϕ ◦ jk , D = ψ ◦ jl . Let m0 ≥ 0 and assume that the sequences of vector bundles (3)

pl+m (ϕ)

pm (ψ)

Jk+l+m (E) −−−−−→ Jl+m (F ) −−−−→ Jm (B)

are exact, for all m ≥ m0 . Then the sequence (2) is formally exact, Rk+l+m is a vector bundle for all m ≥ m0 and Nl+m0 = Ker pm0 (ψ) is a formally integrable differential equation of order l + m0 on F . Moreover, if Rk+l+m0 = Ker pl+m0 (ϕ) is formally integrable, the sequences (4)

σl+m+1 (ϕ)

σm+1 (ψ)

S k+l+m+1 T ∗ ⊗ E −−−−−−−→ S l+m+1 T ∗ ⊗ F −−−−−→ S m+1 T ∗ ⊗ B

are exact for all m ≥ m0 . Proof. The first assertion of the lemma is a direct consequence of Corollary 2, §3, n0 5 of Bourbaki [1965], since finite-dimensional vector spaces are artinian. From Lemma 1.1, it follows that Rk+l+m and Nl+m = Ker pm (ψ) are vector bundles, for m ≥ m0 ; the exactness of the sequences (3) gives us also the surjectivity of πl+m : Nl+m+1 → Nl+m , for m ≥ m0 . For m ≥ m0 , we consider the commutative diagram (5). If Rk+l+m0 is formally integrable, its columns are exact; by means of this diagram, the exactness of the sequences (3) then implies the exactness of (4), for m ≥ m0 . If X is a real-analytic manifold and E is a real-analytic vector bundle over X, we denote by Eω the sheaf of analytic sections of E. Theorem 1.4. Suppose that X is a real-analytic manifold, and that E, F , B are real-analytic vector bundles over X. Let ϕ : Jk (E) → F , ψ : Jl (F ) → B be realanalytic morphisms of vector bundles over X, and set D = ϕ ◦ jk , D = ψ ◦ jl . Let m0 ≥ 0; if the sequences (3) are exact for all m ≥ m0 and Rk+l+m0 = Ker pl+m0 (ϕ) is formally integrable, then the sequence D

D

Eω −→ Fω −→ Bω


361

is exact. Proof. Let f be an analytic section of F over a neighborhood U of x ∈ X satisfying D f = 0. From the exactness of the sequences (3), we deduce that, for m ≥ l + m0 , Nk+m = Kerjm (f) pm (ϕ)|U is an affine sub-bundle of the vector bundle Jk+m (E)|U whose associated vector bundle is Rk+m |U . By means of the commutativity of diagram (5) and the exactness of the sequence (4), with m ≥ m0 , given by Lemma 1.3, we easily see that πk+l+m : Nk+l+m+1 → Nk+l+m is surjective for all m ≥ m0 . According to Theorem 2.2, Chapter IX, there exists an analytic section s of E over a neighborhood of x such that jk+l+m0 (s) is a section of Nk+l+m0 ; then Ds = f on this neighborhood. Goldschmidt [1968a] shows that one can replace the hypothesis “Rk+l+m0 is formally integrable” in the above theorem by the weaker condition: πm : Rm+r → Rm has locally constant rank for all m ≥ k + l + m0 , r ≥ 0. The following example shows that these conditions can not be weakened in an essential way.

362



363

Example 1.5. Suppose that X = R with its standard coordinate x, and that E = F is the trivial real line bundle. We identify a section of E with a real-valued function on X. Consider the analytic first-order differential operator D : E → E given by Df = x2

df − f, dx

where f ∈ E. Let ϕ : J1 (E) → E be the morphism of vector bundles satisfying D = ϕ ◦ j1 , and Rl+1 = Ker pl (ϕ). For l ≥ 0, it is easily verified that Rl+1 is a vector bundle of rank 1 on X and that π0 : Rl+1 → E is an isomorphism on X − {0}; therefore R1 is formally integrable on X − {0}. However dim gl+1,x = 1 for x = 0, and so πl : Rl+1 → Rl is not surjective at x = 0 and does not have constant rank on X. On the other hand, the sequence pl (ϕ)

0 → Rl+1 → Jl+1 (E) −−−→ Jl (E) → 0 is exact, for all l ≥ 0; thus the sequences (3) corresponding to the complex (6)

D

E −→ E → 0

are exact for all m ≥ 0, and this complex is formally exact by Lemma 1.3. However the sub-complex (7)

D

Eω −→ Eω → 0

of (6) is not exact. Indeed, we see that R∞,x = 0 for x = 0. Hence there exists a unique formal solution of infinite order at 0 of the equation (8)

Df = −x;

in fact, if f is a solution of the equation (8) on a neighborhood of 0, then the Taylor series of f at 0 is ∞ n! xn+1 . n=0

Therefore there does not exist a real-analytic function f on a neighborhood of 0 satisfying (8) near 0, and (7) is not exact. One can also see that the complex (6) itself is not exact. We now give the version of Theorem 2.16, Chapter IX for linear equations (Goldschmidt [1967a], Theorem 4.1).

364



365

Theorem 1.6. Let ϕ : Jk (E) → F be a morphism of vector bundles. If (A) Rk+1 is a vector bundle, (B) πk : Rk+1 → Rk is surjective, (C) gk is 2-acyclic, then Rk is a formally integrable linear differential equation. Proof. According to (A), the morphism p1 (ϕ) is of locally constant rank and so its cokernel B is a vector bundle. We denote by ψ : J1 (F ) → B the natural projection; we set p−1 (ψ) = 0. For l ≥ 0, we consider the commutative diagram (9). Its columns are exact, and its rows are complexes and are exact at S k+l+1 T ∗ ⊗ E, Jk+l+1 (E) and Jk+l (E). We denote by hl the cohomology of the top row at S l+1 T ∗ ⊗ F . If l = 0, it follows from (B) and the definition of ψ that the top row is exact, i.e. h0 = 0. Lemma 1.7. Let l ≥ 0. If hl = 0, then we have an isomorphism (10)

hl+1 → H k+l,2 (gk ).

Proof. According to the commutativity of the diagram (23) of Chapter IX, the 62 ∗ T ⊗ diagram (11) commutes and is exact, except perhaps for its first column at gk+l and its first row at S l+2 T ∗ ⊗ F ; hence it gives us a natural isomorphism (10). We now return to the proof of Theorem 1.6. From diagram (9) with l = 1, we obtain an exact sequence πk+1

Ω

Rk+2 −−−→ Rk+1 − → h1 ; if p ∈ Rk+1 , then Ω(p) is the cohomology class of −1 pl+1 (ϕ)q in h1 , where q ∈ Jk+2 (E) satisfies πk+1 q = p; by means of the isomorphism (10) with l = 0, we therefore have an exact sequence πk+1

κ

Rk+2 −−−→ Rk+1 − → H k,2 (gk ). Since H k,2 (gk ) = 0, we see that πk+1 : Rk+2 → Rk+1 is surjective. More generally, by induction on l ≥ 1, from diagram (9), Lemma 1.7 and (C), we simultaneously obtain the surjectivity of πk+l : Rk+l+1 → Rk+l and the exactness of the second row of (9). Moreover, by Lemma 1.1 and the exactness of the top row of (9), gk+l is a vector bundle for l ≥ 1; the exactness of the sequences

πk+l

0 → gk+l+1 − → Rk+l+1 −−−→ Rk+l → 0 and (A) now imply that Rk+l is a vector bundle for all l ≥ 0.

366



367

We thus have completed the proof of Theorem 1.6 and, if we denote by D : F → B the first-order differential operator ψ ◦ j1 , in the process, by Lemma 1.3, we have also proved the following Theorem 1.8. Let ϕ : Jk (E) → F be a morphism of vector bundles and let D be the differential operator ϕ ◦ jk of order k. Assume that conditions (A), (B) and (C) of Theorem 1.6 hold. Then there exist a vector bundle B and a first-order linear differential operator D : F → B such that the sequence D

D

E −→ F −→ B is formally exact; moreover the sequences pl+1 (D)

pl (D )

Jk+l+1 (E) −−−−−→ Jl+1 (F ) −−−−→ Jl (B) are exact, for all l ≥ 0. Again, as in Theorems 2.16 and 3.1 of Chapter IX, according to Theorem 2.14, Chapter IX, we may replace condition (C) by: (C ) for all x ∈ X, there exists a quasi-regular basis of Tx for gk at x. Similarly, using Theorem 2.13, Chapter IX, we have the corresponding version of Theorem 2.17, Chapter IX, which is given by Goldschmidt [1968a], Theorem 4.2. The following generalization of Theorem 1.8 is given by Goldschmidt [1968a], Theorem 3 and [1967a], Theorem 4.4. Theorem 1.9. Let ϕ : Jk (E) → F be a morphism of vector bundles and let D = ϕ ◦ jk . (i) Assume that X is connected and that there is an integer l0 ≥ 0 such that Rk+l is a vector bundle for all l ≥ l0 . Then there exists a complex (12)

D

P

Pj

P

Pj+1

1 2 E −−→ B0 −−→ B1 −−→ B2 → · · · → Bj−1 −−→ Bj −−→ · · · ,

where Bj is a vector bundle and B0 = F , and where Pj : Bj−1 → Bj is a linear differential operator of order lj , which is formally exact; moreover, if r0 = 0 and rj = l1 + l2 + · · · + lj , for j ≥ 1, the sequences pm (D)

pm−r1 (P1 )

pm−r2 (P2 )

Jk+m (E) −−−→ Jm (B0 ) −−−−−−→ Jm−r1 (B1 ) −−−−−−→ Jm−r2 (B2 ) → · · · (13)

pm−rj (Pj )

→ Jm−rj−1 (Bj−1 ) −−−−−−→ Jm−rj (Bj ) → · · ·

are exact at Jm−rj (Bj ) for m ≥ rj+1 and j ≥ 0. (ii) Let 0 = r0 < r1 < · · · < rj < rj+1 < · · · be integers such that Rk+r1 −1 is formally integrable and H k+rj +m−j,j+1 (gk ) = 0 for all j ≥ 1 and m ≥ 0. Then there exists a complex (12), where Bj is a vector bundle and B0 = F , and where Pj : Bj−1 → Bj is a linear differential operator of

368


order lj = rj − rj−1 , for j ≥ 1, which is formally exact; moreover the sequences (13) are exact at Jm−rj (Bj ) for m ≥ rj+1 and j ≥ 0. If the hypotheses of Theorem 1.9,(i) are satisfied and if the mappings πk+l : Rk+l+1 → Rk+l are of constant rank for all l ≥ l0 , the cohomologies of two different sequences (12) are isomorphic and depend, up to isomorphisms, only on Rk+l0 (see Goldschmidt [1968a]). The sequences of type (12) were first introduced by Kuranishi [1964]. We now give a brief outline of the proof of Theorem 1.9,(ii). Since pr1 (ϕ) has locally constant rank, its cokernel B1 is a vector bundle. We denote by ψ1 : Jr1 (F ) → B1 the natural projection and set P1 = ψ1 ◦ jr1 . As Rk+r1 −1 is formally integrable and H k+r1 −1+m,2 (gk ) = 0, for all m ≥ 0, an argument similar to the one given above to prove the exactness of the sequences of vector bundles of Theorem 1.8 tells us that the sequences pr1 +m (ϕ)

pm (ψ1 )

Jk+r1 +m (E) −−−−−−→ Jr1 +m (F ) −−−−→ Jm (B1 ) are exact for all m ≥ 0. If j ≥ 2 and Pi , Bi are defined for 1 ≤ i ≤ j − 1 and if the sequences (13) are exact at Jm−ri (Bi ) for m ≥ ri+1 and 0 ≤ i ≤ j − 1, then, according to Lemma 1.1, plj (Pj−1 ) has locally constant rank and so its cokernel Bj is a vector bundle; we denote by ψj : Jlj (Bj−1 ) → Bj the natural projection and set Pj = ψj ◦ jlj . If pm = lj + m + 1, qm = pm + lj−1 , one shows that the sequences σp

σm+1 (Pj )

(Pj−1 )

m S qm T ∗ ⊗ Bj−2 −−− −−−−→ S pm T ∗ ⊗ Bj−1 −−−−−−→ S m+1 T ∗ ⊗ Bj

are exact for m ≥ 0. Then one deduces that the sequences plj +m (Pj−1 )

pm (Pj )

Jlj−1 +lj +m (Bj−2 ) −−−−−−−−→ Jlj +m (Bj−1 ) −−−−→ Jm (Bj ) are exact for all m ≥ 0. Example 1.10. Let l be an integer ≥ 2 for which Rk+l−1 is formally integrable and gk+l−1 is involutive; then the hypotheses of Theorem 1.9,(ii) hold with rj = l + (j − 1), for j ≥ 1. In this case, one can give a more explicit description of a sequence (12) obtained from Theorem 1.9,(ii) and the above construction. First, let B1 be a vector bundle isomorphic to the cokernel of pl−1 (ϕ) and P1 be a differential operator of order l − 1 such that the sequence pl−1 (D)

p(P )

1 → B1 → 0 Jk+l−1 (E) −−−−−→ Jl−1 (F ) −−−−

is exact. Since gk+l−1 is involutive, one defines vector bundles B1 , . . . , Bn and morphisms of vector bundles σ1 : S l T ∗ ⊗ F → B1 ,

σj : T ∗ ⊗ Bj−1 → Bj ,

for 2 ≤ j ≤ n, as follows. Let B1 be a vector bundle isomorphic to the cokernel of σl (ϕ) and σ1 a morphism such that the sequence σl (ϕ)

1 B1 → 0 S k+l T ∗ ⊗ E −−−→ S l T ∗ ⊗ F −→

σ


369

is exact. We set B0 = B0 = F ; if j ≥ 2, we obtain Bj and σj from Bj−2 , Bj−1 and σj−1 by letting Bj be a vector bundle isomorphic to the cokernel of (σj−1 )+1 and σj be a morphism such that the sequence (σj−1 )+1

σj

S lj−1 +1 T ∗ ⊗ Bj−2 −−−−−−→ T ∗ ⊗ Bj−1 −→ Bj → 0

is exact, where l1 = l and lj = 1 for 2 ≤ j ≤ n. In fact, if we write σjr = (σj )+r , the sequences σm−1

σm

σl+m (ϕ)

1 → S m T ∗ ⊗ B1 −2−→ S m−1 T ∗ ⊗ B2 → S k+l+m T ∗ ⊗ E −−−−→ S l+m T ∗ ⊗F −

(14)

σm−n+1

n −− −−→ S m−n+1 T ∗ ⊗ Bn → 0 · · · → S m−n+2 T ∗ ⊗ Bn−1

are exact for all m ≥ 0. We set Bn+1 = 0 and

6j−1 ∗ T ⊗ B1 ), Bj = Bj ⊕ ( for 1 ≤ j ≤ n + 1. We identify T ∗ ⊗ Bj with (T ∗ ⊗ Bj ) ⊕ (T ∗ ⊗ Let

µj : T ∗ ⊗

6j−2

6j−1

T ∗ ⊗ B1 →

T ∗ ⊗ B1 ).

6j−1

T ∗ ⊗ B1

be the morphism sending α ⊗ ω ⊗ u into (α ∧ ω) ⊗ u, for α ∈ T ∗ , ω ∈ u ∈ B1 . For 2 ≤ j ≤ n + 1, we write

6j−2

T ∗ and

νj = (σj , µj ) : T ∗ ⊗ Bj−1 → Bj . Then there exists a differential operator P1 : F → B1 of order l such that σ(P1 ) = σ1 and P1 · D = 0, and we define P1 : F → B1 by P1 f = (P1 f, P1 f), for f ∈ F. If 2 ≤ j ≤ n + 1 and Pj−1 is defined, then Pj : Bj−1 → Bj is the unique first-order differential operator such that σ(Pj ) = νj and Pj · Pj−1 = 0. We obtain a sequence (15)

D

P

Pn+1

P

1 2 E −−→ B0 −−→ B1 −−→ B2 → · · · → Bn −−→ Bn+1 → 0

such that the sequences pl+m (D)

pm (P1 )

pm−1 (P2 )

Jk+l+m (E) −−−−→ Jl+m (B0 ) −−−→ Jm (B1 ) −−−−−→ Jm−1 (B2 ) → · · · (16)

pm−n (Pn+1 )

→ Jm−n+1 (Bn ) −−−−−−−→ Jm−n (Bn+1 ) → 0 are exact for all m > 0. According to Lemma 1.3, the equation Ker p(Pj ) obtained from the sequence (12) given by Theorem 1.9,(ii) is formally integrable, for j ≥ 1. If X is a realanalytic manifold, E, F are real-analytic vector bundles and ϕ : Jk (E) → F is an

370


analytic morphism of vector bundles satisfying the hypotheses of Theorem 1.9,(ii), according to the above construction of the sequence (12), the vector bundles Bj and the operators Pj given by Theorem 1.9,(ii) can be chosen to be analytic. Then by Theorem 1.4, the sub-complex D

P

Pj

P

Pj+1

1 2 (B1 )ω −−→ (B2 )ω → · · · → (Bj−1 )ω −−→ (Bj )ω −−→ · · · Eω −−→ (B0 )ω −−→

of (12) is exact. If ϕ is analytic and satisfies the hypotheses of Theorem 1.9,(i), and if moreover πm : Rm+r → Rm has constant rank for all m ≥ k + l0 and r ≥ 0, then the vector bundles Bj and the operators Pj given by Theorem 1.9,(i) can also be chosen to be analytic, and the above sub-complex of (12) is exact (see Goldschmidt [1968a], Corollary 4). Goldschmidt [1970b] proves that, if X is connected and ϕ is analytic, then, outside an analytic set, ϕ satisfies all these regularity conditions: Rk+l is a vector bundle for all l ≥ 0, and πm : Rm+r → Rm has constant rank for all m ≥ k and r ≥ 0. If ϕ satisfies the hypotheses of Theorem 1.9,(ii) together with additional assumptions, the proof of Theorem 1.9,(ii) gives us the existence of a sequence of type (12) whose vector bundles can be explicitly described. Theorem 1.11. Let ϕ : Jk (E) → F be a morphism of vector bundles and let D = ϕ ◦ jk . Assume that Rk = Ker ϕ is formally integrable and that σ(D) : S k T ∗ ⊗ E → F is surjective. Let q ≥ 0 be an integer. (i) Let 0 = r0 < r1 < · · · < rq = rq+1 be integers such that the cohomology groups H k+rj−1 +m−j−1,j+1 (gk )

(17)

vanish for all 1 ≤ j ≤ q + 1 and m ≥ 1, with m = rj − rj−1 . Then there exists a complex (18)

D

P

Pq

P

1 2 B1 −−→ B2 → · · · → Bq−1 −−→ Bq → 0, E −−→ B0 −−→

where Bj is a vector bundle isomorphic to H k+rj −j−1,j+1(gk ) for j ≥ 0 and B0 = F , and where Pj is a linear differential operator of order lj = rj − rj−1 for 1 ≤ j ≤ q, such that the sequences σm−l (P1 )

σm (D)

σm−l

−l

(P2 )

S k+m T ∗ ⊗ E −−−→ S m T ∗ ⊗ B0 −−−−1−−→ S m−l1 T ∗ ⊗ B1 −−−−1−−2−−→ S m−l1 −l2 T ∗ ⊗ B2 → · · · → S m−l1 −···−lq−1 T ∗ ⊗ Bq−1

(19)

σm−l1 −···−lq (Pq )

−−−−−−−−−−→ S m−l1 −···−lq T ∗ ⊗ Bq → 0 are exact for all m ≥ 0. (ii) Let (18) be a complex, where Bj is a vector bundle, B0 = F and Pj is a linear differential operator of order lj ≥ 1. If the sequences (19) are exact for all m ≥ 0, the complex (18) is formally exact; moreover the sequences pm (D)

pm−l (P1 )

pm−l

−l

(P2 )

1 −→ Jm−l1 (B1 ) −−−−1−−2−−→ Jm−l1 −l2 (B2 ) → Jk+m (E) −−−→ Jm (F ) −−−−

· · · → Jm−l1 −···−lq (Bq ) → 0


371

are exact for all m ≥ 0. Furthermore, if we set r0 = 0 and rj = l1 + · · · + lj , for 1 ≤ j ≤ q, and rj = rq for j > q, then the cohomology groups (17) vanish for all j ≥ 1 and m ≥ 0 except possibly for m = lj and j = 1, . . . , q, and Bj is isomorphic to H k+rj −j−1,j+1 (gk ). æ Example 1.12. If k = 1 and g1 is involutive, and if R1 = Ker ϕ is formally integrable and σ(D) : T ∗ ⊗ E → F is surjective, then the hypotheses of Theorem 1.11,(i) hold, with q = n − 1 and rj = j for 0 ≤ j ≤ n − 1. We thus obtain a complex (18), where Bj is equal to 6 6 H 0,j+1 (g1 ) = ( j+1 T ∗ ⊗ E)/δ( j T ∗ ⊗ g1 ), for 1 ≤ j ≤ n − 1, for which the sequences (19) are exact for all m ≥ 0. In fact, since δ(α ∧ u) = (−1)i α ∧ δu, 6i ∗ 6j ∗ for α ∈ T and u ∈ T ⊗ S m T ∗ ⊗ E, the morphism (20)

T∗ ⊗

6j

T∗ ⊗ E →

6j+1

T ∗ ⊗ E,

sending α ⊗ ω ⊗ u into (α ∧ ω) ⊗ u, for α ∈ T ∗ , ω ∈ passage to the quotient a morphism of vector bundles

6j

T ∗ , u ∈ E, induces by

σj : T ∗ ⊗ H 0,j (g1 ) → H 0,j+1(g1 ). Since σ(D) is surjective, the diagram 0 ⏐ ⏐ "

0 ⏐ ⏐ "

0 ⏐ ⏐ " σ(D)

F −−−−→ 0 ⏐ ⏐ "id

p(D)

F −−−−→ 0 ⏐ ⏐ "

0 −−−−→ g1 −−−−→ T ∗ ⊗ E −−−−→ ⏐ ⏐ ⏐ ⏐ " " 0 −−−−→ R1 −−−−→ J1 (E) −−−−→ ⏐ ⏐ ⏐π ⏐π " 0 " 0 0 −−−−→ E −−−−→

E ⏐ ⏐ "

−−−−→ 0

0 is commutative and exact; therefore π0 : R1 → E is surjective and we may identify B0 = F with H 0,1 (g1 ) in such a way that σ(D) = σ0 . Then, if we write P0 = D, the differential operators Pj : Bj−1 → Bj are uniquely determined by the relations σ(Pj ) = σj and Pj · Pj−1 = 0, for 1 ≤ j ≤ n − 1 (see Goldschmidt [1967a], §5). The resulting complex (18) is called the sophisticated Spencer sequence of the first-order equation R1 .

372


Example 1.13. If there exists an integer l ≥ 0 such that Rk+l = Ker pl (ϕ) is formally integrable and gk+l+1 is involutive, the sophisticated Spencer sequence of Rk+l is a complex D

D

D

0 1 2 C 0 −−→ C 1 −−→ C 2 −−→ · · · → C n → 0,

(21)

where C 0 = Rk+l and C j is a vector bundle over X isomorphic to the cokernel of the composition 6j−1

T ∗ ⊗ gk+l+1 − → δ

6j

id ⊗

T ∗ ⊗ gk+l −−→

6j

T ∗ ⊗ Rk+l ,

and where Dj is a first-order linear differential operator such that: (i) Ker p(D0 ) = λ1 Rk+l+1 is formally integrable, its symbol is involutive, and the mapping jk+l : E → Jk+l (E) induces, by restriction, an isomorphism {u ∈ E | Du = 0} → {v ∈ C 0 | D0 v = 0}; (ii) the sequences σm−1 (D0 )

σm−2 (D1 )

S m T ∗ ⊗ C 0 −−−−−→ S m−1 T ∗ ⊗ C 1 −−−−−→ S m−2 T ∗ ⊗ C 2 → · · · → S m−n T ∗ ⊗ C n → 0 are exact for all m ≥ 0; (iii) if (12) is a complex, where B0 = F and where Pj is a differential operator of order lj and rj = l1 + · · · + lj , for which the sequences (13) are exact at Jm−rj (Bj ) for m ≥ rj+1 and j ≥ 0, its cohomology at Bj−1 is isomorphic to the cohomology of (21) at C j . The sophisticated Spencer sequences were originally introduced by Spencer [1962]; other constructions are given in Bott [1963], Quillen [1964], Goldschmidt [1967a] and Spencer [1969]. Under certain regularity assumptions on the differential operator D, Theorem 1.14,(i) gives us the existence of a compatibility condition D : F → B for D and the following theorem (see Goldschmidt [1968a], Theorem 1 and [1968b], Theorem 2) tells us that the solvability questions for the complex D

D

E −→ F −→ B can be reduced to those pertaining to a formally exact complex D

D

1 1 E −−→ F1 −−→ B1

for which Ker p(D1 ) is formally integrable and D1 = P · D, where P : F → F1 is a differential operator obtained from D in finitely many steps. In particular, if f is a section of F , the solutions of the inhomogeneous equation Du = f are the same as those of the equation D1 u = P f.


373

Theorem 1.14. Let ϕ : Jk (E) → F be a morphism of vector bundles and let D = ϕ◦jk . Assume that X is connected and that there is an integer r0 ≥ 0 such that Rk+l is a vector bundle for all l ≥ r0 and that the mappings πk+l : Rk+l+r → Rk+l have constant rank for all l ≥ r0 and r ≥ 0. Then there exist integers m0 ≥ r0 and l0 ≥ 0, a vector bundle F1 and a differential operator P : F → F1 of order m0 + l0 such that the following assertions hold: (i) The differential operator D1 = P · D : E → F1 is of order k+m0 and, if ψ = p(D1 ) : Jk+m0 (E) → F1 , the equation Ker ψ is equal to πk+m0 Rk+m0 +l0 and is formally integrable; the solutions and formal solutions of this equation are exactly those of Rk+r0 . (ii) Let B, B1 be vector bundles and D : F → B, D1 : F1 → B1 be differential operators of order l and q respectively. If the sequences pr (D )

pl+r (D)

Jk+l+r (E) −−−−−→ Jl+r (F ) −−−−→ Jr (B), pr (D )

pq+r (D1 )

Jk+m0 +q+r (E) −−−−−−→ Jq+r (F1 ) −−−−1→ Jr (B1 ) are exact for all r ≥ 0, there exist an integer m ≥ 0 and a differential operator Q : B → Jm (B1 ) satisfying the following conditions: (a) the diagram D

E −−−−→ ⏐ ⏐ "id

D

F −−−−→ ⏐ ⏐ "P

B ⏐ ⏐Q "

jm ◦D

D

1 1 E −−−− → F1 −−−−→ Jm (B1 )

commutes. Hence, if f ∈ F satisfies D f = 0, then D1 (P f) = 0. (b) For f ∈ F satisfying D f = 0, the solutions u ∈ E of the inhomogeneous equation Du = f coincide with the solutions of the equation D1 u = P f. Moreover, P induces an isomorphism from the cohomology of the complex D

D

E −→ F −→ B to the cohomology of the complex D

D

1 1 E −−→ F1 −−→ B1 .

Hence, if f1 ∈ F1 satisfies D1 f1 = 0, there exists f ∈ F satisfying P f = f1 and D f = 0. The first part of the above theorem is the prolongation theorem of Goldschmidt [1968a]; a result generalizing it together with the Cartan–Kuranishi prolongation theorem is given by Goldschmidt [1974], Theorem 1.

374


§2. Examples. Example 2.1. Let E be a vector bundle over X. A connection on E is a linear differential operator ∇:E →T∗⊗E such that ∇(fs) = df ⊗ s + f∇s,

(22)

where f is a real-valued function on X and s is a section of E. According to (1), the relation (22) is equivalent to the fact that ∇ is a first-order differential operator whose symbol σ(∇) : T ∗ ⊗ E → T ∗ ⊗ E is the identity mapping. Let ∇ be a connection on E. It determines a splitting χ∇ : E → J1 (E) of the exact sequence of vector bundles 0 0 → T∗ ⊗ E − → J1 (E) −→ E→0

π

satisfying ◦ p(∇) = id − χ∇ ◦ π0 and a first-order differential equation R1 = Ker p(∇) = χ∇ (E); clearly π0 : R1 → E is an isomorphism and the symbol of R1 is equal to 0. By (22), the first-order linear differential operator d∇ :

6j

T∗ ⊗E →

6j+1

T∗⊗E

determined by d∇ (ω ⊗ s) = dω ⊗ s + (−1)j ω ∧ ∇s,

(23) for ω ∈ (24)

6j

T ∗ and s ∈ E, is well-defined. According to (23), its symbol σ(d∇ ) : T ∗ ⊗

6j

T∗ ⊗ E →

6j+1

T∗ ⊗ E

is equal to (20). Moreover, the differential operator d∇ · ∇ : E →

62

T∗⊗E

is of order zero and arises from a morphism of vector bundles K:E→

62

T ∗ ⊗ E,

the curvature of ∇, which we shall identify with a section of fact, we have K(ξ, η)s = (∇ξ ∇η − ∇η ∇ξ − ∇[ξ,η] )s,

62

T ∗ ⊗ E ∗ ⊗ E; in

§2. Examples

375

for ξ, η ∈ T , s ∈ E. We denote by Rl+1 the l-th prolongation of R1 . Using Proposition 2.8, Chapter IX, it can be shown that the sequence π

K◦π

1 0 R1 −−−→ R2 −→

62

T∗ ⊗ E

is exact; since the symbol of R1 is equal to 0, by Theorem 1.6, we therefore see that R1 is formally integrable if and only if the curvature K of ∇ vanishes (see Gasqui and Goldschmidt [1983], Theorem 2.1). Suppose that the curvature K of ∇ vanishes. Then, according to (23), the sequence (25)

d∇

∇

E −→ T ∗ ⊗ E −→

62

d∇

T ∗ ⊗ E −→ · · · →

6n

T∗⊗E →0

is a complex. Since the morphism (24) is equal to (20), and (25) is a complex, the construction given in Example 1.12 shows that (25) is the sophisticated Spencer sequence of R1 . Let U be a connected and simply-connected open subset of X; it is well-known that the mapping (26)

{s ∈ C ∞ (U, E) | ∇s = 0} → Ex ,

sending s into s(x), is an isomorphism of vector spaces. Let E be the trivial vector bundle U × Ex and ϕ : E|U → E be the unique isomorphism of vector bundles sending s ∈ C ∞ (U, E) satisfying ∇s = 0 into the constant section a → (a, s(x)) of E over U . Then the diagram 6i

T∗⊗E ⏐ ⏐ ∇ "d

6i+1

id ⊗ϕ

−−−−→

id ⊗ϕ

T ∗ ⊗ E −−−−→

6i

T ⊗ E ⏐ ⏐ "d

6i+1

T ∗ ⊗ E

over U , where d is the exterior differential operator acting on functions with values in Ex , commutes. Therefore by the Poincaré lemma, the sequence (25) is exact and we have proved all but the last assertion of the next theorem, which follows from the fact that H 1 (U, R) = 0 (see Gasqui and Goldschmidt [1983], Proposition 2.1). Theorem 2.2. Let ∇ be a connection on E. Then the curvature K of ∇ vanishes if and only if R1 = Ker p(∇) is formally integrable. If K = 0, then the complex (25) is the sophisticated Spencer sequence of R1 and is exact; moreover, if U is a connected and simply-connected open subset of X, the mapping (26) is an isomorphism and the sequence ∇

d∇

C ∞ (U, E) −→ C ∞ (U, T ∗ ⊗ E) −−→ C ∞ (U,

62

T ∗ ⊗ E)

is exact. Example 2.3. Let V be a finite-dimensional vector space and let E be the trivial vector bundle X ×V . The exterior differential operator d : E → T ∗ ⊗E for V -valued

376


functions on X is a connection on E. The sequence (25) corresponding to d is the de Rham sequence (27)

E− → T∗ ⊗E − → d

d

62

T∗⊗E − → ··· → d

6n

T ∗ ⊗ E → 0,

which is exact by the Poincaré lemma. Thus the curvature of d vanishes and (27) is therefore the sophisticated Spencer sequence of the formally integrable equation on E equal to Ker p(d). Example 2.4. We denote by FC the complexification of a real vector bundle F . 6k E denote the Assume that E is a complex vector bundle over X. Let S k E and k-th symmetric and exterior powers (over C) of the complex vector bundle E. The bundle Jk (E) is a complex vector bundle if we set c · jk (s)(x) = jk (c · s)(x), for x ∈ X, s ∈ Ex and c ∈ C. The real vector bundle S k T ∗ ⊗ E is canonically isomorphic to the complex vector bundle S k TC∗ ⊗C E and we shall identify these two vector bundles. The morphism : S k TC∗ ⊗C E → Jk (E) sends ((df1 · . . . · dfk ) ⊗ s)(x) into jk

(7

k i=1 fi

) · s (x),

where f1 , . . . , fk are complex-valued functions on X vanishing at x and s is a section of E over X. If F is also a complex vector bundle and D : E → F is a C-linear differential operator of order k, then pl (D) : Jk+l (E) → Jl (F ) is a morphism of complex vector bundles, as is the morphism σl (D) : S k+l TC∗ ⊗C E → S l TC∗ ⊗C F. Assume now that n = 2m + k, where m ≥ 1 and k ≥ 0. An almost CR-structure (of codimension k) on X is a complex sub-bundle E of TC of rank m (over C) such that E and its complex conjugate E have a zero intersection. This almost CR-structure is said to be a CR-structure if E is stable under the Lie bracket, i.e. [E , E ] ⊂ E . Let OX be the sheaf of complex-valued functions on X and let 1C denote the trivial complex line bundle over X. If E is an almost CR-structure on X, let ρ : TC∗ → E ∗ be the projection induced by the inclusion E → TC and let ∂ b : OX → E ∗ be the first-order differential operator which is equal to the composition of the exterior differential operator d : OX → TC∗ and ρ : TC∗ → E ∗ . The symbol σ(∂ b ) : TC∗ → E ∗ of ∂ b is equal to ρ and so is surjective; therefore R1 = Ker p(∂ b )

§2. Examples

377

is a first-order differential equation on 1C and π0 : R1 → 1C is surjective (see Example 1.12). The symbol g1 of R1 is equal to the annihilator E ⊥ of E . The k-th prolongation gk+1 of g1 is equal to the sub-bundle S k+1 E ⊥ of S k+1 TC∗ . Let x ∈ X; a basis {t1 , . . . , tn } of TC,x over C, for which the subspace of TC spanned (over C) by the vectors tj , with n − m + 1 ≤ j ≤ n, is equal to Ex , is quasi-regular for the sub-bundle g1 = E ⊥ of TC∗ at x in the sense that dimC g2,x = dimC g1,x +

n−1

∗ dimC (TC,x,{t ∩ Ex ⊥ ), 1 ,...,tj }

j=1 ∗ ∗ is the annihilator in TC,x of the subspace of TC,x spanned by where TC,x,{t 1 ,...,tj } {t1 , . . . , tj } over C. Then the arguments proving that condition (i) of Theorem 2.14, Chapter IX implies condition (ii) show that g1 is involutive. Moreover, it is easily seen that 6 6 H 0,j (g1 ) = j TC∗ /(E ⊥ ∧ j−1 TC∗ ). 6j ∗ 6j ∗ TC → E induces an isomorphism of vector Since the restriction mapping ρ : bundles

6j

(28)

TC∗ /(E ⊥ ∧

6j−1

TC∗ ) →

6j

E ∗ ,

we obtain a canonical isomorphism ψj : H 0,j (g1 ) →

6j

E ∗ .

Clearly, the diagram σj

(29)

T ∗ ⊗ H 0,j (g1 ) −−−−→ H 0,j+1 (g1 ) ⏐ ⏐ ⏐id ⊗ψ ⏐ψ j " " j+1 6 6 γ j j j+1 ∗ T ∗ ⊗ E ∗ −−−−→ E

is commutative, where σj is defined in Example 1.12 and γj is the morphism TC∗ ⊗C 6j+1 ∗ 6j ∗ E → E sending α ⊗ β into ρ(α) ∧ β. It is easily seen that the sequence σ1 (∂ b )

γ1

0 → S 2 E ⊥ → S 2 TC∗ −−−−→ TC∗ ⊗C E ∗ −→ is exact. We now compute the first obstruction Ω : R1 → of R1 . If p ∈ R1 , then Ω(p) = γ1 −1 p1 (∂ b )q,

62

62

E ∗ → 0

E ∗ to the integrability

where q ∈ J2 (1C ) satisfies π1 q = p. According to Proposition 2.8, Chapter IX, the sequence (30) is exact.

π

Ω

1 R2 −→ R1 − →

62

E ∗

378


Lemma 2.5. Let x ∈ X. If f ∈ OX,x satisfies (∂ b f)(x) = 0 and u is the element −1 j1 (∂ b f)(x) of TC∗ ⊗C E , then (31)

u(ζ, η(x)) = ζ · (η · f),

(32)

Ω(j1 (f)(x))(ξ(x), η(x)) = [ξ, η], df(x),

for all ξ, η ∈ Ex , ζ ∈ TC,x . Proof. We can write ∂bf =

r

gj α j ,

j=1

where αj ∈ Ex , gj ∈ OX,x and gj (x) = 0. Then u=

r

(dgj ⊗ αj )(x)

j=1

and u(ζ, η(x)) = (ζ · gj )η, αj (x). On the other hand, η · f = η, ∂ b f =

r

η, gj αj =

j=1

and so ζ · (η · f) =

r

r

gj η, αj

j=1

(ζ · gj )η, αj (x),

j=1

since gj (x) = 0. Formula (32) is a direct consequence of (31). Proposition 2.6. Let E be an almost CR-structure on X. Then the following four statements are equivalent: (i) E is a CR-structure; 62 ∗ E such that the diagram (ii) there exists a differential operator ∂ b : E ∗ → d

62

∂

62

TC∗ −−−−→ ⏐ ⏐ρ " b E ∗ −−−− →

TC∗ ⏐ ⏐ρ " E ∗

commutes; (iii) π1 : R2 → R1 is surjective; (iv) R1 is a formally integrable differential equation. Proof. The existence of the operator ∂ b of (ii) is easily seen to be equivalent to the following condition: if α ∈ E ⊥ , then (dα)(ξ, η) = 0, for all ξ, η ∈ E . From the formula (dα)(ξ, η) = ξ · α(η) − η · α(ξ) − α([ξ, η]),

§2. Examples

379

we deduce the equivalence of (i) and (ii) (see Kuranishi [1977], Proposition 1). From the exactness of sequence (30), we see that (iii) is equivalent to the vanishing of Ω. If f is a constant function on X, then ∂ b f = 0 and Ω(j1 (f)(x)) = 0, for x ∈ X. Therefore (iii) holds if and only if the mapping Ω ◦ : E ⊥ →

62

E ∗

is equal to zero. By (32), we have (Ω ◦ )(α)(ξ, η) = [ξ, η], α, for ξ, η ∈ E , α ∈ E ⊥ . Thus Ω ◦ = 0 if and only if E is a CR-structure, and so (i) is equivalent to (iii). Since g1 = E ⊥ is an involutive sub-bundle of TC∗ and g2 is a vector bundle, the equivalence of (iii) and (iv) is provided by Theorem 1.6. From the isomorphisms (28) and Proposition 2.6, it follows that, if E is a CRstructure, for all j ≥ 0 there exists a differential operator ∂b : such that the diagram

6j

E ∗ →

6j+1

E ∗

6j

d

6j+1

6j

∂

6j+1

TC∗ −−−−→ ⏐ ⏐ρ " E ∗ −−−b−→

TC∗ ⏐ ⏐ρ " E ∗

commutes. Clearly, b b → E ∗ − → OX −

∂

(33)

∂

62

b E ∗ − → ··· →

∂

6m

E ∗ → 0

is a complex and ∂ b (α ∧ β) = ∂ b α ∧ β + (−1)j α ∧ ∂ b β,

(34) for all α ∈

6j

E ∗ , β ∈

6r

E ∗ . According to (34) and (1),

σ(∂ b ) : TC∗ ⊗C

6j

E ∗ →

6j+1

E ∗

is equal to γj . Since σ(∂ b ) : TC∗ → E ∗ is surjective and g1 is involutive, from the isomorphisms ψj and the commutativity of diagram (29), if E is a CR-structure, we see that (33) is isomorphic to the sophisticated Spencer sequence of the formally integrable first-order equation R1 , described in Example 1.12. Example 2.7. We continue to use the notation and terminology introduced in Example 2.4. If X is a complex manifold, the sub-bundle T of TC of tangent vectors of type (0, 1) satisfies the conditions T ∩ T = 0 and [T , T ] ⊂ T . The complex structure of X determined by T is a CR-structure (of codimension zero), and in this case the operator ∂ b is equal to the Cauchy–Riemann operator ∂ : OX → T ∗

380


and, if n = 2m, the sequence (33) is the Dolbeault sequence (35)

OX − → T ∗ − → ∂

∂

62

T ∗ − → ··· → ∂

6m

T ∗ → 0

of X, which is always exact (see §3). More generally, let X be a real submanifold of a complex manifold Y of codimension k. If i : X → Y is the inclusion mapping and TY is the bundle of tangent vectors to Y of type (0, 1), let E be the sub-bundle of TC with variable fiber determined by , i∗ (Ex ) = i∗ (TC,x ) ∩ TY,i(x)

for x ∈ X. If E is a vector bundle, then clearly it is a CR-structure of codimension k on X; if k = 1, it is easily seen that this condition always holds. If E is a vector bundle, the operator ∂ b : OX → E ∗ is called the tangential Cauchy–Riemann operator; its solutions include the restrictions to X of the holomorphic functions on Y . The complex (33) was first introduced in this case, with k = 1, by Kohn and Rossi [1965]. Example 2.8. In the preceding example, let Y = C2 , with complex coordinates (z, w) and let X be the real hypersurface of Y defined by the equation Im w = |z|2 . We consider the induced CR-structure E (of codimension 1) on X and its tangential Cauchy–Riemann operator ∂ b , which is essentially the famous locally non-solvable example of H. Lewy [1957]: the sequence OX −→ E ∗ → 0 ∂b

(36) given by (33) is not exact.

We now again consider the morphism of vector bundles ϕ : Jk (E) → F and the objects associated to it. If Rk+l is a vector bundle and there exists an integer m ≥ 0 such that gk+l+m = 0, we say that Rk+l is a differential equation of finite type. In this case, from Theorem 2.2 and the proof of Theorem 2.2 of Gasqui and Goldschmidt [1983], we obtain: Theorem 2.9. Let ϕ : Jk (E) → F be a morphism of vector bundles and let D be the differential operator ϕ ◦ jk of order k. Let l ≥ 0 be an integer such that Rk+l = Ker pl (ϕ) is formally integrable and gk+l+1 = 0. Let (12) be a complex, where Bj is a vector bundle and Pj is a linear differential operator of order lj . If rj = l1 + · · · + lj , suppose that the sequences (13) are exact at Jm−rj (Bj ) for m ≥ rj+1 and j ≥ 0. Then the complex (12) is exact and moreover, if U is a connected and simply-connected open subset of X and x ∈ U , the mapping {s ∈ C ∞ (U, E) | Ds = 0} → Rk+l,x , sending s into jk+l (s)(x), is an isomorphism of vector spaces and the sequence 1 C ∞ (U, B1 ) C ∞ (U, E) −→ C ∞ (U, F ) −→

D

P

is exact. To prove Theorem 2.9, one shows that the cohomology of the sequence (12) is isomorphic to the cohomology of the complex (25), where E = Rk+l and ∇ is the

§2. Examples

381

connection on Rk+l corresponding to the unique morphism χ∇ : Rk+l → J1 (Rk+l ) for which the diagram id Rk+l+1 −−−−→ Rk+l+1 ⏐ ⏐ ⏐πk+l ⏐ " "λ1 Rk+l

χ

−−−∇−→ J1 (Rk+l )

commutes, where πk+l is an isomorphism. Using Theorem 2.2, one sees that the curvature of ∇ vanishes and that the desired assertions hold. 5q ∗ T , we denote by Lξ β the Lie derivative of Example 2.10. If β is a section of β along a vector field ξ on X. Let g be a Riemannian metric on X. We wish to describe the compatibility condition for the first-order linear differential operator D : T → S2 T ∗, sending ξ into Lξ g, under various assumptions on g. We consider the morphism ϕ = p(D), with E = T and F = S 2 T ∗ and some of the objects introduced in Example 2.3, Chapter IX. The mapping σ(ϕ) : T ∗ ⊗ T → S 2 T ∗ is surjective and the fiber g1,x of its kernel g1 at x ∈ X is equal to the Lie algebra of the orthogonal group of the Euclidean vector space (Tx , g(x)); moreover g2 = 0 (see Gasqui and Goldschmidt [1983], §3). Thus we see that R1 = Ker p(ϕ) is a differential equation of finite type. The solutions of the equation R1 or of the homogeneous equation Dξ = 0 are the Killing vector fields of (X, g). We denote by Ej the sub-bundle of 6j+1 ∗ 62 ∗ T ⊗ T , which is the kernel of the morphism of vector bundles µ:

6j+1

T∗ ⊗

62

T∗ →

6j+2

T∗ ⊗ T∗

determined by for ω ∈ by:

6j+1

µ(ω ⊗ (α ∧ β)) = (α ∧ ω) ⊗ β − (β ∧ ω) ⊗ α, T ∗ , α, β ∈ T ∗ ; then E1 = G. The Spencer cohomology of g1 is given H 0,0 (g1 ) = T,

(37)

H 0,1 (g1 ) S 2 T ∗ ,

H 1,0 (g1 ) = H 1,1 (g1 ) = 0,

H 0,j (g1 ) = 0,

H 1,j (g1 ) Ej−1 ,

for j > 1, and H m,i (g1 ) = 0, for m ≥ 2, i ≥ 0, where the isomorphisms depend only on g (see Gasqui and Goldschmidt [1983], §3). Let H be the sub-bundle of T ∗ ⊗ G consisting of those elements ω of T ∗ ⊗ G which satisfy the relation ω(ξ1 , ξ2 , ξ3 , ξ4 , ξ5 ) + ω(ξ2 , ξ3 , ξ1 , ξ4 , ξ5 ) + ω(ξ3 , ξ1 , ξ2 , ξ4 , ξ5 ) = 0, for all ξ1 , ξ2 , ξ3 , ξ4 , ξ5 ∈ T . In fact, according to the second Bianchi identity (DR)(g) = ∇g R(g) is a section of H. Let δ : H → T ∗ ⊗ G be the inclusion mapping; the image Bj of the morphism δ:

6j−1

T∗ ⊗ H →

6j

T ∗ ⊗ G,

382


6j−1 ∗ sending ω ⊗ u into (−1)j−1 ω ∧ δu, for ω ∈ T , u ∈ H, is a sub-bundle of 6j ∗ T ⊗ G. According to the exactness of the sequences (3.25l) and (3.31l) of Gasqui and Goldschmidt [1983], there exist morphisms σ1 : S 3 T ∗ ⊗ S 2 T ∗ → H and σj : T ∗ ⊗ Bj−1 → Bj , for 2 ≤ j ≤ n, such that the sequences (14) are exact for all m ≥ 0; in fact, σ1 is the first prolongation of τ . We set ∇ = ∇g and R = R(g). Let Rg : S 2 T ∗ → G,

(DR)g : S 2 T ∗ → H

be the linear differential operators of order 2 and 3 which are the linearizations along g of the non-linear operators h → R(h) and h → (DR)(h) respectively, where h is a Riemannian metric on X. If h ∈ S 2 T ∗ , we therefore have d d (DR)g (h) = (DR)(g + th)|t=0 . Rg (h) = R(g + th)|t=0 , dt dt The invariance of the operators R and DR gives us the formulas (38)

Rg (Lξ g) = Lξ R,

(DR)g (Lξ g) = Lξ ∇R,

for all ξ ∈ T (see Gasqui and Goldschmidt [1983], Lemma 4.4). By formula (4.30) of Gasqui and Goldschmidt [1983], we have σ((DR)g ) = σ1 . ˜ be the sub-bundle of G with variable fiber, whose fiber at x ∈ X is Let G ˜ x = {(Lξ R)(x) | ξ ∈ Tx with (Lξ g)(x) = 0}, G ˜ be the canonical projection. If G ˜ is a vector bundle, we define and let α : G → G/G a second-order linear differential operator D1 : S 2 T ∗ → G/G˜ by setting

(D1 h)(x) = α(Rg (h − Lξ g))(x),

for x ∈ X and h ∈ S 2 Tx∗ , where ξ is an element of Tx satisfying h(x) = (Lξ g)(x) whose existence is guaranteed by the surjectivity of σ(D). By the first relation of (38), we see that this operator is well-defined; clearly we have D1 · D = 0. By Proposition 5.1 of Gasqui and Goldschmidt [1983], the sequence p2 (D)

p(D1 )

˜→0 J3 (T ) −−−−→ J2 (S 2 T ∗ ) −−−→ G/G is exact. For j ≥ 1, we consider the vector bundle 6 ˜ Bj = Bj ⊕ ( j−1 T ∗ ⊗ G/G) and we let

P1 : S 2 T ∗ → B1

be the third-order linear differential operator given by P1 h = ((DR)g (h), D1 h), for h ∈ S 2 T ∗ .

5q ∗ For x ∈ X, let ρ denote the representation of g1,x on Tx and also on Tx . We say that (X, g) is a locally symmetric space if ∇R = 0. According to Theorem 7.1 and §5 of Gasqui and Goldschmidt [1983], we have: æ

§2. Examples

383

Theorem 2.11. If (X, g) is a locally symmetric space, R3 = Ker p2 (ϕ) is a ˜ is a vector bundle equal to the “informally integrable differential equation and G finitesimal orbit of the curvature” {ρ(u)R | u ∈ g1 }. We now suppose that ∇R = 0. From the second relation of (38), we deduce that P1 · D = 0. Since g2 = 0, the hypotheses of Theorem 1.9,(ii) hold for ϕ with k = 1, rj = j + 2, for j ≥ 1. Therefore, if νj : T ∗ ⊗ Bj−1 → Bj is the morphism defined in Example 1.10, for 2 ≤ j ≤ n + 1, according to the construction given there, there exists a complex (39)

Pn+1

1 2 T −−→ S 2 T ∗ −−→ B1 −−→ B2 → · · · → Bn −−→ Bn+1 → 0,

P

D

P

with σ(Pj ) = νj , for 2 ≤ j ≤ n + 1, of the type (15) of Example 1.10 such that the sequences (16) are exact for all m ≥ 0, with E = T , B0 = S 2 T ∗ . In Gasqui and Goldschmidt [1983], §7, the operators P2 , . . . , Pn+1 are determined explicitly (see also Gasqui and Goldschmidt [1988a]). From Theorem 2.9, we now deduce the following result of Gasqui and Goldschmidt [1983] (Theorem 7.2): Theorem 2.12. If ∇R = 0, the complex (39) is exact and, if U is a simplyconnected open subset of X, the sequence P1 D ˜ C ∞ (U, T ) −→ C ∞ (U, S 2 T ∗ ) −→ C ∞ (U, H ⊕ G/G)

is exact. According to Lemma 5.3 of Gasqui and Goldschmidt [1983], the assumption ˜ =0 H ∩ (T ∗ ⊗ G) implies that (X, g) is a locally symmetric space; if this condition is satisfied, there is an exact sequence Pn+2

P

2 1 G/G˜ −−→ B2 → · · · → Bn+1 −−→ Bn+2 →0 T −−→ S 2 T ∗ −−→

D

D

obtained from the complex of type (15) corresponding to the differential operator D1 (see Gasqui and Goldschmidt [1988a], Theorem 2.3). Proposition 6.1 of Gasqui and Goldschmidt [1983] tells us that Proposition 2.13. If X is connected, the differential equation R1 is formally integrable if and only if (X, g) has constant curvature. 6l ∗ Consider the connection ∇ in T and the corresponding differential operator d∇ :

6j+1

T∗⊗

6l

T∗→

6j+2

T∗⊗

6l

T∗

of Example 2.1; according to Gasqui and Goldschmidt [1983], §6, if l = 2, it induces by restriction a first-order linear differential operator d∇ : Ej → Ej+1 ,

for j ≥ 1.

384


Suppose that (X, g) has constant curvature K; then R = −Kτ (g ⊗ g). Consider the first-order linear differential operator Dg : S 2 T ∗ → G defined by

Dg h = Rg (h) + 2Kτ (h ⊗ g),

for h ∈ S 2 T ∗ . According to Gasqui and Goldschmidt [1983], §6 and [1984a], §16, the Calabi sequence (40)

d∇

Dg

d∇

T −−→ S 2 T ∗ −−→ E1 −−→ E2 −−→ · · · → En−1 → 0 D

is a complex. By (37), the hypotheses of Theorem 1.11,(i) are satisfied with k = 1, q = n − 1, rj = j + 1, for 1 ≤ j ≤ n − 1; it is shown in Gasqui and Goldschmidt [1983] that (40) is a complex of the type (18), for which the sequences (19) are exact. From Theorem 2.9, we obtain: Theorem 2.14. If (X, g) has constant curvature, the complex (40) is exact and, if U is a simply-connected open subset of X, the sequence Dg

C ∞ (U, T ) −→ C ∞ (U, S 2 T ∗ ) −−→ C ∞ (U, G) D

is exact. The sequence (40) is the resolution of the sheaf of Killing vector fields of the space of constant curvature (X, g) introduced by Calabi [1961]. The sequences of Example 2.10 have been used to study infinitesimal rigidity questions for symmetric spaces related to the Blaschke conjecture, namely by Bourguignon for the real projective spaces RPn (see Besse [1978]), and by Gasqui and Goldschmidt [1983, 1984b, 1988b, 1989a, 1989b] for the complex projective spaces CPn , with n ≥ 2, the complex quadrics of dimension ≥ 5, and for arbitrary products of these spaces with flat tori. In particular, the infinitesimal orbit of the curvature is determined explicitly for these spaces. Example 2.15. Let g be a Riemannian metric on X. We say that the Riemannian manifold (X, g) is conformally flat if, for every x ∈ X, there is a diffeomorphism ϕ of a neighborhood U of x onto an open subset of Rn and a real-valued function u on U such that ϕ∗ g = eu g, where g is the Euclidean metric on Rn . We consider some of the objects introduced in Example 2.10. Let Trj :

6j+1

Tr : S 2 T ∗ → R, T∗ ⊗ T∗ ⊗ T∗ →

6j

T∗ ⊗ T∗

be the trace mappings defined by Tr h = (Trj u)(ξ1 , . . . , ξj , η) =

r i=1 n i=1

h(ti , ti ), u(ti , ξ1 , . . . , ξj , ti , η),

§2. Examples

385

6j+1 ∗ for all h ∈ S 2 Tx∗ , u ∈ ( T ⊗ T ∗ ⊗ T ∗ )x , ξ1 , . . . , ξj , η ∈ Tx , with x ∈ X, where {t1 , . . . , tn } is an orthonormal basis of Tx . We denote by S02 T ∗ the kernel of Tr and we set Ej0 = Ej ∩ Ker Trj . If ( , ) is the scalar product on T ∗ induced by g, we consider the scalar product 6l ∗ ( , ) on T determined by (41)

(α1 ∧ · · · ∧ αl , β1 ∧ · · · ∧ βl ) = det (αi , βj ),

for α1 , . . . , αl , β1 , . . . , βl ∈ T ∗ . This scalar product in turn induces a scalar product 6j+1 ∗ 62 ∗ T ⊗ T and hence on its sub-bundle Ej . We denote by ρj : Ej → Ej0 on the orthogonal projection onto Ej0 . We suppose that n ≥ 3; we wish to determine the compatibility condition for the first-order linear differential operator D : T → S02 T ∗ sending ξ into 12 Lξ g − n1 Tr(Lξ g)g , when (X, g) is conformally flat. We consider the morphism ϕ = p(D) with E = T and F = S02 T ∗ . The mapping σ(ϕ) : T ∗ ⊗T → S02 T ∗ is surjective. We have g3 = 0 and so R1 = Ker p(D) is a differential equation of finite type; the solutions of R1 or of the homogeneous equation Dξ = 0 are the conformal Killing vector fields of (X, g). The only non-zero Spencer cohomology groups of g1 are given by: H 0,0 (g1 ) = T, (42)

H 0,1 (g1 ) S02 T ∗ ,

0 , for 2 ≤ j ≤ n − 2, H 1,j (g1 ) Ej−1 6 6 n ∗ 2,n−1 2 ∗ H (g1 ) T ⊗ S0 T , H 2,n (g1 ) = n T ∗ ⊗ T ∗ ,

where the isomorphisms depend only on g (see Gasqui and Goldschmidt [1984a], §2). The Weyl tensor W(g) of (X, g) is the section ρ1 R of E10 . If n ≥ 4, a result of H. Weyl asserts that the Riemannian manifold (X, g) is conformally flat if and only if its Weyl tensor vanishes. If n = 3, we have E10 = 0 and the Weyl tensor always vanishes. Proposition 5.1 of Gasqui and Goldschmidt [1984a] tells us that Proposition 2.16. If n ≥ 4, the differential equation R1 is formally integrable if and only if W = 0. The linearization along g of the non-linear differential operator h → W(h), where h is a Riemannian metric on X, is the second-order linear differential operator Wg : S 2 T ∗ → E1 defined by d W(g + th)|t=0 . dt If W = W(g) = 0, the operator Wg takes its values in E10 . We consider the firstorder differential operators Wg (h) =

0 Pj+1 = ρj+1 d∇ : Ej0 → Ej+1 ,

386


for 1 ≤ j ≤ n − 4, and Pn−1 = d∇ · Trn−1 : Let φ:

6n−2

6n

T ∗ ⊗ S02 T ∗ →

T∗ ⊗ T∗ ⊗ T∗ ⊗ T∗ →

6n−1

6n

T ∗ ⊗ T ∗. 62

T∗ ⊗

T∗

be the morphism sending β1 ⊗β2 ⊗α1 ⊗α2 into (β1 ∧α1 )⊗(β2 ∧α2 ), for β1 ∈ β2 , α1 , α2 ∈ T ∗ . The mapping φˆ :

6n−2

T∗ ⊗ T∗ →

6n−1

T∗ ⊗

62

6n−2

T ∗,

T∗

sending u into φ(u ⊗ g) is an isomorphism. Let R0 be the section of T ∗ ⊗ T determined by g(R0 (ξ), η) = Ric(ξ, η), where Ric is the Ricci curvature of g, and ξ, η ∈ T . We define a morphism θ:

6n−2

T∗ ⊗

62

T∗ →

6n−1

T∗ ⊗ T∗

by θ(u)(ξ1 , . . . , ξn−1, η) =

n−1 1 (−1)l+n−1 u(ξ1 , . . . , ξˆl , . . . , ξn−1 , R0 (ξl ), η), n−2 l=1

6n−2 ∗ 62 ∗ for u ∈ T ⊗ T , ξ1 , . . . , ξn−1, η ∈ T . Assume that n ≥ 4 and that W = 0. Then there exists a unique second-order differential operator 6 0 Pn−2 : En−3 → n T ∗ ⊗ S02 T ∗ such that 6 (−1) 0 (d∇ · φˆ−1 · d∇ − θ) : En−3 → n−1 T ∗ ⊗ T ∗ 2 n

Trn−1 Pn−2 =

(see Gasqui and Goldschmidt [1984a], Chapter I). By (42), the hypotheses of Theorem 1.11,(i) are satisfied with k = 1, q = n − 1, rj = j + 1, for 1 ≤ j ≤ n − 3, and rn−2 = n, rn−1 = n + 1. Gasqui and Goldschmidt [1984a] show that the sequence Wg

Pn−2

2 0 (43) T − → S02 T ∗ − → E10 − → · · · → En−3 −−→

D

P

6n

Pn−1

T ∗ ⊗ S02 T ∗ −−→

6n

T∗ ⊗T∗ → 0

is a complex of type (18), for which the sequences (19) are exact. From Theorem 2.9, we obtain the following result of Gasqui and Goldschmidt [1984a]: Theorem 2.17. If n ≥ 4 and (X, g) is conformally flat, the sequence (43) is exact and, if U is a simply-connected open subset of X, the sequence Wg

C ∞ (U, T ) −→ C ∞ (U, S02 T ∗ ) −−→ C ∞ (U, E10 ) D

§3. Existence Theorems for Elliptic Equations

387

is exact. Gasqui and Goldschmidt [1984a] also construct a resolution of type (18) of the sheaf of conformal Killing vector fields of a conformally flat space when n = 3. We now return to the morphism ϕ of §1, the corresponding differential operator D : E → F and the objects associated to them. If U is an open subset of X, let C0∞ (U, E) denote the space of sections of E with compact support contained in U . Suppose that E and F are endowed with scalar products ( , ). The formal adjoint D∗ : F → E of D is the unique differential operator such that (Du, v) = (u, D∗ v), U

U

for any oriented subset U of X and for all u ∈ C0∞ (U, E), v ∈ C ∞ (U, F ); it is of order k. Let (12) be a complex of differential operators for which the sequences (13) are exact at Jm−rj (Bj ) for m ≥ rj+1 and j ≥ 0. If the vector bundles E, F , Bj are endowed with scalar products, in general the adjoint complex Pj∗

∗ Pj−1

P∗

P∗

D∗

2 1 · · · → Bj −−→ Bj−1 −−→ · · · → B2 −−→ B1 −−→ B0 − → E,

with B0 = F , does not necessarily satisfy the analogous condition. However, there are just a few examples for which it does, namely: Example 2.3 (continued). Let g be a Riemannian metric on X and fix a scalar product on the vector space V . Then we obtain scalar products on the vector 6 bundles j T ∗ ⊗ E; the adjoint sequence 6n

d∗

T ∗ ⊗ E −→

6n−1

d∗

T ∗ ⊗ E −→ · · · →

62

d∗

d∗

T ∗ ⊗ E −→ T ∗ ⊗ E −→ E → 0

of (27) is the sophisticated Spencer sequence of the equation Ker p(d∗ ), where 6n ∗ 6n−1 ∗ T ⊗E → T ⊗ E, and is exact. d∗ : Example 2.15 (continued). The vector bundles of the sequence (43) all inherit scalar products from the metric g. If n ≥ 4 and W = 0, the adjoint sequence ∗ ∗ 6 Pn−1 Pn−2 6n ∗ P1∗ D∗ 0 T ⊗ T ∗ −−→ n T ∗ ⊗ S02 T ∗ −−→ En−3 → · · · → E10 − → S02 T ∗ − →T∗ →0 of (43), where P1 = Wg , is again a complex of type (18), for which the sequences (19) are exact; moreover it is exact (see Gasqui and Goldschmidt [1984a]). §3. Existence Theorems for Elliptic Equations. We again consider the morphism ϕ : Jk (E) → F , the differential operator D = ϕ ◦ jk : E → F and the objects of §1 associated to them. If x ∈ X and α ∈ Tx∗ , let σα (D) : Ex → Fx be the linear mapping defined by σα (D)u =

1 σ(D)(αk ⊗ u), k!

388


where u ∈ Ex and αk denotes the k-th symmetric product of α. If E and F are endowed with scalar products and D∗ : F → E is the formal adjoint of D, then we have (44)

σα (D∗ ) = (−1)k σα (D)∗ .

If B is another vector bundle and D : F → B is a differential operator of order l, then, for all x ∈ X and α ∈ Tx∗ , it is easily verified that (45)

σα (D · D) = σα (D ) · σα (D) : Ex → Bx .

We say that α ∈ Tx∗ , with x ∈ X, is non-characteristic for D if the mapping σα (D) : Ex → Fx is injective. Thus α is non-characteristic for D if and only if there are no non-zero elements e of Ex such that αk ⊗ e ∈ gk . Assume now that k = 1 and thus that D is a first-order operator. We say that a subspace U of Tx∗ , with x ∈ X, is non-characteristic for D if (U ⊗ E) ∩ g1 = 0. Let U be a sub-bundle of T ∗ ; if x0 ∈ X and g1,x0 is involutive and if Ux0 is a maximal non-characteristic subspace of Tx∗0 for D, then, for all x ∈ X belonging to a neighborhood of x0 , the subspace Ux of Tx∗ is maximal non-characteristic. Assume that D is a first-order differential operator satisfying the following conditions: (i) R1 = Ker p(D) is formally integrable; (ii) σ(D) : T ∗ ⊗ E → F is surjective; (iii) g1 is involutive. We consider the initial part (46)

D

D

E −→ F −→ B

of the sophisticated Spencer sequence of the first-order equation R1 given by Example 1.12, where D is a differential operator of order 1. We now describe the normal forms for D and for the complex (46) introduced by Guillemin [1968] (see also Spencer [1969]). Given a point x0 ∈ X, let Ux0 be a maximal non-characteristic subspace of Tx∗0 . We choose a local coordinate system (x1 , . . . , xn ) for X on a neighborhood of x0 such that dx1, . . . , dxk forms a basis for Ux0 at x0 . The subspace Ux of Tx∗ spanned by dx1 , . . . , dxk at x is maximal non-characteristic for D for all x belonging to a neighborhood V of x0 . Let U be the sub-bundle of T ∗ |V spanned by the sections dx1 , . . . , dxk . We consider the complex (46) restricted to V . For i = 1, . . . , k, the morphism σdxi (D) : E → F is injective and the vector bundle Ei = σdxi (D)(E) is isomorphic to E. It is easily seen that the sum E1 + · · · + Ek is direct, and we choose a complement E0 to this sum in F ; thus we have F = E0 ⊕ E1 ⊕ · · · ⊕ Ek .


389

For i = 0, 1, . . . , k, we let πi : F → Ei be the projection and we set Di = σdxi (D)−1 · πi · D,

D0 = π0 D,

for i = 1, . . . , k. Then in terms of the differential operators D0 : E → E0 , we have D = D0 +

Di : E → E,

k

σdxi (D)Di .

i=1

¿From the definitions, for i = 1, . . . , k, we see that ∂ Di = + Li , ∂xi and that ∂ ∂ ∂ ∂ Li = Li D0 = D0 , . . ., n , , . . ., n ∂xk+1 ∂x ∂xk+1 ∂x are differential operators involving differentiations only along the sheets of the foliation x1 = constant, x2 = constant, . . . , xk = constant. The main results of Guillemin [1968] are summarized in Theorem 3.1. For 1 ≤ i, j ≤ k, there exist first-order differential operators Dij : E0 → E,

such that

[Di , Dj ] = Dij D0 ,

Di : E0 → E0 D0 Di = Di D0 ,

where [Di , Dj ] = Di Dj −Dj Di . Moreover, the differential operator D0 also satisfies conditions (i), (ii) and (iii); if D0 : E0 → B0 is the first-order differential operator which is the compatibility condition for D0 whose symbol σ(D0 ) : T ∗ ⊗ E0 → B0 is surjective, the complex (46) can be described as follows: We may identify F with E0 ⊕ (U ⊗ E) and B with 62 B0 ⊕ (U ⊗ E0 ) ⊕ ( U ⊗ E) and the operators D and D are given by Du = D0 u +

k

dxi ⊗ Di u,

i=1

D (f0 +

k

dxi ⊗ fi )

i=1

= D0 f0 +

k

dxi ⊗ (D0 fi − Di f0 )

i=1

+

dxi ∧ dxj ⊗ (Di fj − Dj fi − Dij f0 ),

1≤i<j≤k

for u ∈ E, f0 ∈ E0 , f1 , . . . , fk ∈ E. We no longer make any assumptions on the operator D. We say that D is a determined (resp. an underdetermined) differential operator if, for all x ∈ X, there exists a non-zero cotangent vector α ∈ Tx∗ such that σα (D) is an isomorphism (resp. is surjective). The following result is due to Quillen [1964] (see also Gasqui [1976]).

390


Proposition 3.2. Let ϕ : Jk (E) → F be a morphism of vector bundles and let D = ϕ ◦ jk . Then the differential operator D is underdetermined if and only if the morphisms σl (ϕ) : S k+l T ∗ ⊗ E → S l T ∗ ⊗ F are surjective for all l ≥ 0. If D is underdetermined, then Rk = Ker ϕ is formally integrable and the morphisms pl (ϕ) : Jk+l (E) → Jl (F ) are surjective for all l ≥ 0. ¿From the proof of Theorem 1.6, it follows that the second assertion of this proposition is an easy consequence of the first. From Proposition 3.2, we infer that, if D is underdetermined, then g1 is involutive. We say that D is elliptic (resp. determined elliptic, underdetermined elliptic) if for all x ∈ X and α ∈ Tx∗ , with x ∈ X, the mapping σα (D) is injective (resp. is an isomorphism, is surjective). If there exists an integer l ≥ 0 such that gk+l = 0, then D is elliptic (see Goldschmidt [1967a], Proposition 6.2). Lemma 3.3. Let B be a vector bundle over X and D : F → B be a differential operator of order k. Assume that E, F , B are endowed with scalar products. If for all x ∈ X and α ∈ Tx∗ , with α = 0, the sequence σα (D)

σα (D )

Ex −−−−→ Fx −−−−→ Bx

(47)

is exact, then the differential operator = DD∗ + D∗ D : F → F of order 2k is determined elliptic. Proof. Let x ∈ X and α ∈ Tx∗ , with α = 0. If u ∈ Ex, by (44) and (45), we have (σα ()u, u) = (−1)k ((σα (D)σα (D)∗ + σα (D )∗ σα (D ))u, u) = (−1)k {(σα (D)∗ u, σα(D)∗ u) + (σα (D )u, σα(D )u)}. If σα ()u = 0, we deduce that σα (D )u = 0 and σα (D)∗ u = 0. From the exactness of the sequence (47), we infer that u = 0 and so is determined elliptic. Theorem 3.4. If D : E → F is a determined elliptic or underdetermined elliptic operator, then the sequence D E −→ F → 0 is exact. If D is determined elliptic, this result is classic (see Hörmander [1963], Theorem 7.5.1). If D is underdetermined elliptic, one chooses scalar products on the vector bundles E and F ; by Lemma 3.3, the operator DD∗ : F → F is determined elliptic and the theorem for D follows from the previous case. Let (48)

Q0

Q1

Q2

Qr−1

B0 −−→ B1 −−→ B2 −−→ · · · → Br−1 −−→ Br → 0


391

be a complex, where Bj is a vector bundle over X and Qj is a linear differential operator of order lj ≥ 1. We say that (48) is an elliptic complex if, for all x ∈ X and α ∈ Tx∗ , with α = 0, the sequence σα (Q0 )

σα (Q1 )

σα (Qr−1 )

0 → B0,x −−−→ B1,x −−−→ B2,x → · · · → Br−1,x −−−−−→ Br,x → 0 is exact. Examples 2.7 and 2.8 (continued). If X is a complex manifold, the Cauchy–Riemann operator ∂ : OX → T ∗ is elliptic and the Dolbeault sequence (35) is an elliptic complex. On the other hand, if E is a CR-structure on X of codimension > 0, the tangential Cauchy–Riemann operator ∂ b : OX → E ∗ fails to be elliptic. The operator ∂ b of Example 2.8 is determined. The following theorem is a direct consequence of the proof of Quillen’s theorem (see Quillen [1964] and Goldschmidt [1967a], Proposition 6.5; see also Proposition 6.2, Chapter VIII). Theorem 3.5. Suppose that the complex (48) satisfies the following condition: for all m ≥ 0, the sequences σm−l (Q1 )

σm (Q0 )

σm−l

−l

(Q2 )

S m+l0 T ∗ ⊗ B0 −−−−→ S m T ∗ ⊗ B1 −−−−1−−→ S m−l1 T ∗ ⊗ B2 −−−−1−−2−−→ · · · σm−l1 −···−lr−1 (Qr−1 )

→ S m−l1 −···−lr−2 T ∗ ⊗ Br−1 −−−−−−−−−−−−−→ S m−l1 −···−lr−1 T ∗ ⊗ Br → 0 are exact. If Q0 is elliptic, then (48) is an elliptic complex. If D is elliptic and if there exists an integer l ≥ 0 such that Rk+l is formally integrable and gk+l+1 is involutive, then, by Proposition 6.4 of Goldschmidt [1967a], the operator D0 of the sophisticated Spencer sequence (21) of Rk+l is also elliptic, and so from Theorem 3.5 we deduce that (21) is an elliptic complex. We are interested in knowing under which conditions an elliptic complex is exact. Proposition 3.2 asserts that an elliptic complex (48) with r = 1 is a complex of type (12) and Theorem 3.4 tells us that it is always exact. If r > 1, the following example shows that a formally exact elliptic complex is not necessarily exact, and hence that in this case we must restrict our attention to complexes of type (12). √ Example 3.6. Suppose that X = C with its complex coordinate z = x + −1 y. We set √ √ ∂ ∂ ∂ 1 ∂ ∂ 1 ∂ = − −1 , = + −1 . ∂z 2 ∂x ∂y ∂z 2 ∂x ∂y Let E be the trivial complex line bundle over X and F = E ⊕ E. Consider the first-order differential operator D : E → F given by Df =

2 ∂f

∂f z − f, ∂z ∂z

,

for f ∈ E. Let B = E and D : F → B be the first-order differential operator defined by ∂v ∂u D (u, v) = − z2 + v, ∂z ∂z

392


for u, v ∈ E. Then we obtain a complex D

D

E −→ F −→ B → 0.

(49)

∂ Since σα ∂z : Ex → Ex is an isomorphism, for x ∈ X and α ∈ Tx∗ , with α = 0, we easily see that (49) is an elliptic complex. If Rl+1 = Ker pl (D), the sequence pl (D)

pl−1 (D )

0 → Rl+1 → Jl+1 (E) −−−→ Jl (F ) −−−−−→ Jl−1 (B) → 0 is exact for l ≥ 0, and the sequence (49) is formally exact by Lemma 1.3. Indeed, since ∂ : Jl+1 (E) → Jl (E) pl ∂z is surjective for all l ≥ 0, we see that pl−1 (D ) is surjective for l ≥ 1. Moreover, let a ∈ X and u, v ∈ Ea satisfy pl−1 (D )(jl (u)(a), jl (v)(a)) = 0; ( ) there exists f1 ∈ Ea such that jl ∂f ∂z (a) = jl (v)(a). Set

(50)

u = u − z 2

∂f1 + f1 ; ∂z

( ) (a) = 0. Then according to the complex because of (50), we obtain jl−1 ∂u ∂z analytic analogue of Example 1.5, we can choose a holomorphic function f2 on a neighborhood of a such that ∂f2 − f2 − u (a) = 0; jl z 2 ∂z if f = f1 + f2 , we have

jl (Df − (u, v))(a) = 0.

Therefore Rl+1 is a vector bundle for all l ≥ 0. As in Example 1.5, it is easily seen that π0 : Rl+1 → E is an isomorphism on X − {0} and hence that R1 is formally integrable on X − {0}. However, dimC gl+1,a = 1 for a = 0, and so πl : Rl+1 → Rl is not surjective at a = 0 and does not have constant rank on X. On the other hand, the complex (49) is not exact. In fact, consider (−z, 0) ∈ C ∞ (F ); we have D (−z, 0) = 0 and a solution of the equation Df = (−z, 0) is a holomorphic function f satisfying z 2 ∂f ∂z −f = −z. According to the argument given in Example 1.5, there does not exist such a function f on a neighborhood of 0. We no longer impose any of the above restrictions on the operator D. Assume that X is connected and that there is an integer r0 ≥ 0 such that Rk+l is a vector bundle for all l ≥ r0 and that the mappings πk+l : Rk+l+r → Rk+l have constant rank for all l ≥ r0 and r ≥ 0. Then according to Theorem 1.9,(i), we may consider a complex (12), where Bj is a vector bundle over X and B0 = F , and where Pj is a differential operator of order lj ; if rj = l1 + · · · + lj , we may suppose that the sequences (13) are exact at Jm−rj (Bj ) for m ≥ rj+1 and j ≥ 0. If D is not elliptic, in general the complex (12) will not be exact; indeed, we have seen that the complex (36), which is the Spencer sequence associated to the first-order operator ∂ b of Example 2.8, is not exact. We now state


393

The Spencer Conjecture. If D is elliptic, the complex (12) is exact. The main cases when this conjecture is known to be true for overdetermined operators are described in the following two theorems; the first one is due to Spencer and the second one to MacKichan, Sweeney and Rockland. Theorem 3.7. Suppose that X is a real-analytic manifold, E, F are real-analytic and D is an analytic differential operator. If D is elliptic, the complex (12) is exact. Theorem 3.8. Suppose either that E is a real line bundle, or that E is a complex line bundle, F is a complex vector bundle and that D is C-linear. If D is an elliptic first-order operator and if Rl+1 is formally integrable for some l ≥ 0, the complex (12) is exact. The important special case of Theorem 3.8, namely that of a first-order elliptic operator, acting on the trivial complex line bundle and determined by complex vector fields, for which R1 is formally integrable, can be treated using the Newlander–Nirenberg theorem (see Trêves [1981], Theorem 1.1, Chapter I). To verify the Spencer Conjecture, it suffices to consider the case of operators D for which Rk = Ker p(D) is formally integrable. Indeed, according to Goldschmidt [1968a, 1968b], we may replace the differential operator D by the operator D1 = P ·D given by Theorem 1.14,(i); then D1 is elliptic, Ker p(D1 ) is formally integrable and the cohomology of the sequence (12) is isomorphic to that of a sequence (12) corresponding to D1 . We now suppose that Rk = Ker p(D) is formally integrable. Let l ≥ 0 be an integer for which gk+l+1 is involutive. According to Example 1.13, to prove the Spencer Conjecture for D, it suffices to show that the sophisticated Spencer sequence (21) of Rk+l is exact at C j , for j > 0. By Theorem 3.5 and the remark which follows it, if D is elliptic, (21) is an elliptic complex. The first-order operator D0 of (21) satisfies the assumptions considered above for the existence of Guillemin normal forms. We now give a proof due to Spencer [1962] of the exactness of the complex (21) under the hypotheses of Theorem 3.7; it consists of an adaptation of H. Cartan’s argument for the exactness of the Dolbeault sequence (see also Spencer [1969]). As we have seen above, Theorem 3.7 is a consequence of this result. Assume that the hypotheses of Theorem 3.7 hold. Then the vector bundles and the differential operators Dj of (21) are real-analytic. Let x0 be an arbitrary point of X. We choose real-analytic scalar products on the vector bundles C j over a neighborhood U of x0 and consider the complex (21) restricted to U . The Laplacian ∗ + Dj∗ Dj : C j → C j , = Dj−1 Dj−1

with Dj = 0 for j = −1 and n, is real-analytic. As (21) is an elliptic complex, according to Lemma 3.3, is a determined elliptic operator. Now let u be a section of C j over a neighborhood of x0 satisfying Dj u = 0. By Theorem 3.4, there exists a section v of C j over a neighborhood V ⊂ U of x0 such that v = u on V . ∗ We set w = u − Dj−1 Dj−1 v; since Dj u = 0, we have Dj w = 0 on V . On the other hand, ∗ ∗ ∗ Dj−1 w = Dj−1 u − Dj−1 v = 0, and so w = 0 on V . Since is a determined elliptic, analytic operator, it follows that w is real-analytic (see Hörmander [1963], Chapter X). According to

394


Theorem 1.11,(ii) and Theorem 1.4, there exists an analytic section w of C j−1 over a neighborhood V of x0 such that Dj−1 w = w on V ; thus ∗ u = Dj−1 (Dj−1 v + w )

on a neighborhood of x0 , and this proves Theorem 3.7. Example 2.7 (continued). The Cauchy–Riemann operator ∂ : OX → T ∗ of a complex manifold X is real-analytic (with respect to the structure of real-analytic manifold on X determined by its complex structure). Since the first-order equation R1 = Ker p(∂) is formally integrable and involutive and (35) is the Spencer sequence of R1 described in Example 1.12, Theorem 3.7 implies that the sequence (35) is exact. We now describe the δ-estimate discovered by Singer. For the moment, we no longer make any of the above assumptions on E and the operator D. We use some of the notation introduced in Example 2.4, and suppose that E and F are complex vector bundles and that D : E → F is a C-linear differential operator. Hermitian scalar products ( , ) on E and F determine a Hermitian scalar product on E ⊗C F by setting (e1 ⊗ f1 , e2 ⊗ f2 ) = (e1 , e2 )(f1 , f2 ), for e1 , e2 ∈ E, f1 , f2 ∈ F . We fix Hermitian scalar products ( , ) on TC∗ and E. 6 We obtain a Hermitian scalar product ( , ) on l TC∗ for which (41) holds for all 5l ∗ α1 , . . . , αl , β1 , . . . , βl ∈ TC∗ . The imbedding of S l T ∗ into T defined in §1, 5l ∗ l ∗ Chapter IX extends uniquely to a C-linear imbedding S TC → TC ; we identify 5l ∗ S l TC∗ with its image in TC under this mapping. Thus the Hermitian scalar 5l ∗ product on TC induces a scalar product on S l TC∗ , and the Hermitian scalar ∗ products on TC and E induce a scalar product on 6j ∗ 6j ∗ T ⊗ S k+l T ∗ ⊗ E = TC ⊗C S k+l TC∗ ⊗C E. 6j ∗ T ⊗ gk+l We consider the restriction ( , ) of this Hermitian scalar product to 6 and set u2 = (u, u), for u ∈ j T ∗ ⊗ gk+l . Definition. We say that the C-linear differential operator D : E → F of order k satisfies the δ-estimate if there exist Hermitian scalar products on TC∗ and E such that δu2 ≥ u2 , for all u ∈ (T ∗ ⊗ gk+1 ) ∩ Ker δ ∗ , where δ ∗ : T ∗ ⊗ gk+1 → gk+2 is the adjoint of δ. The following result of Rockland [1972] (Appendix A) is verified by adapting the proof of MacKichan [1971] (Example 4.3) that the Cauchy–Riemann operator ∂ : OX → T ∗ of Example 2.7 satisfies the δ-estimate. Proposition 3.9. If E is a complex line bundle, then a first-order C-linear differential operator D : E → F satisfies the δ-estimate. Spencer [1962] proposed a generalization of the ∂-Neumann problem for overdetermined linear elliptic systems on small convex domains. This Spencer–Neumann problem has been studied by Sweeney [1968, 1976] and has been solved only in the case of operators satisfying the δ-estimate by MacKichan [1971], using the work of Kohn and Nirenberg [1965]. An estimate of Sweeney [1969] implies that the corresponding harmonic spaces vanish and gives us the following result due to MacKichan [1971] and Sweeney [1969].


395

Theorem 3.10. Suppose that D is a C-linear elliptic differential operator satisfying the δ-estimate. If there exists an integer l ≥ 0 such that Rk+l is formally integrable, then the complex (12) is exact. Theorem 3.8 now follows from Proposition 3.9 and the above theorem. The δ-estimate also gives a generalization of symmetric hyperbolic operators, in conjunction with the Guillemin normal form of Theorem 3.1 (see MacKichan [1975]). Example 3.11. Let X be an open subset √ of Rk × Cr , with coordinates (t, z) = 1 k 1 r j j (t , . . . , t , z , . . . , z ). We write z = x + −1 yj and √ √ ∂ ∂ 1 ∂ 1 ∂ ∂ ∂ = − −1 j , = + −1 j . ∂z j 2 ∂xj ∂y 2 ∂xj ∂y ∂z j Let E be the trivial complex line bundle of rank m over X and F be the direct sum of (r + 1)-copies of E. Let P1 : E → E,

P2 : E → E

be the first-order differential operators given by P1 u =

k

ai (t, z)

i=1

∂u , ∂ti

P2 u =

r

bj (t, z)

j=1

∂u , ∂z j

for u ∈ E, where aj (t, z), bl (t, z) are m × m matrices which are holomorphic in z. Assume that k ≥ 1 and that P1 is determined in the t-variables, i.e. for all x ∈ X, there exists (51)

α=

k

αj dtj ∈ Tx∗ − {0}

j=1

for which σα (P1 ) : Ex → Ex is an isomorphism. Consider the first-order differential operator D : E → F defined by ∂u ∂u Du = P1 u + P2 u, 1 , . . . , r , ∂z ∂z for u ∈ E. Then using a variant of Proposition 3.2, one verifies that R1 = Ker p(D) is a formally integrable differential equation, that σ(D) : T ∗ ⊗ E → F is surjective r+1 -copies of E, and that g1 is involutive. Moreover, if B is the direct sum of 2 the differential operator D : F → B defined by ∂u0 ∂uq ∂ul D (u0 , u1 , . . . , ur ) = − (P + P )u , − , 1 2 j ∂z q j,l,q=1,...,r ∂z j ∂z l l

Exterior differential systems

Exterior differential systems and Euler-Lagrange PDEs

Exterior Differential Systems (Mathematical Sciences Research Institute Publications)